West Wave · AI · 2026-08-16 · Issue #38

Today's Highlights: Google and Anthropic launched an offense-defense battle around "AI Watermarks" on the same day. Google allows users to turn off watermarks on generated content, while Anthropic thoroughly explained its Claude text watermarking mechanism. 1Password released the SCAM benchmark, specifically testing AI agents' anti-fraud capabilities. Additionally, several enterprise-grade agent tools were open-sourced; AI is moving from "chatting well" to "working reliably."

Editor's Note: The keyword for this week is "Trust." Watermarks, anti-fraud benchmarks, agent identity verification, and security hardening—giants are starting to seriously solve the missing piece of AI implementation: making people dare to entrust important tasks to it.

---

I. Large AI Models

1. Google Gemini Allows Disabling Visible Watermarks on AI-Generated Content

• Summary: Google updated Gemini settings to allow users to remove visible watermarks from AI-generated images, videos, and music. Invisible watermarks (C2PA metadata) remain. Previously, the industry widely questioned whether visible watermarks were an important line of defense for labeling AI content and preventing misuse, and if turning them off would make AI content harder to identify.
• Source: The Verge · 2026-08-15
• Editor's Comment: Google gave users a choice, but controversy followed immediately—one less layer of "AI traces" distinguishable by ordinary users. The watermark debate is essentially a tug-of-war between "creator rights vs. user experience," with no standard answer yet.

2. Anthropic Details Claude Text Watermarking Mechanism, Addressing External Concerns

• Summary: Anthropic published a blog post explaining how watermarks in Claude-generated text will work: embedding statistical patterns at the generation end, without affecting text quality, and providing a detection API for third-party verification. Some users worried about "leaving traces" and chose to unsubscribe; Anthropic emphasized that watermarks can be detected and also opted out of.
• Source: TechCrunch · 2026-08-15
• Editor's Comment: Exactly opposite to Google's direction—Anthropic is building watermarks into trust infrastructure that is "included by default, with public detection." How to label AI content, or whether to label it at all, has put the two giants on opposite ends of the scale.

3. 1Password Releases SCAM Benchmark: Teaching AI Agents Not to Get Scammed

• Summary: 1Password open-sourced the SCAM benchmark, using realistic multi-turn work scenarios to test AI agents' security alertness: phishing emails, fake payment requests, social engineering tactics, etc. The goal is to turn "anti-fraud" into a quantifiable capability for agents, not just verbal reminders.
• Source: 1Password · 2026-08-15
• Editor's Comment: As news spreads that AI can do work for humans, scammers are also studying how to trick AI. Providing "anti-fraud training" for agents is a lesson that must be learned before Agent deployment.

4. AI Thinks the Internet is Reddit: 298 Out of 300 Queries Failed

• Summary: An experiment covering 8,616 responses showed that GPT-4.1, GPT-5.5, and GPT-5.6 cited Reddit content in 298 out of 300 queries where no platform was mentioned, treating Reddit as the primary source of internet information. Smaller models performed more restrainedly.
• Source: Growtika · 2026-08-15
• Editor's Comment: AI's "worldview" is skewed by the density of Reddit in its training corpus. This reminds everyone using AI for information gathering: its "common sense" is likely just the average opinion of a specific community.

II. AI Software

1. AstraZeneca Releases R&D Agent System: Research Assistant

• Summary: AstraZeneca revealed its internal agent system used for the entire R&D process: literature search, experimental design, and data interpretation are all assisted by agents. The paper claims this is an important sample of a pharmaceutical giant applying Agentic AI to real R&D pipelines.
• Source: arXiv · 2026-08-15
• Editor's Comment: Pharma is among the first industries daring to put AI on the "production line." A research assistant won't solve scientific discovery, but freeing scientists from literature and data is already tangible efficiency.

2. Atuin AI Server Open-Sourced: Bringing Terminal Agents Back Home

• Summary: Terminal tool Atuin announced the open-sourcing of its AI Server, allowing users to self-host a terminal agent that understands their work habits—it can browse history, understand command lines, and run repetitive operations, keeping data completely local.
• Source: Atuin · 2026-08-16
• Editor's Comment: The terminal is a programmer's last private territory. Putting AI in the terminal while self-hosting means wanting intelligence without surrendering data. These "local-first" tools will become increasingly common.

3. Colleague's View: Writing Code with AI Feels More Like Being a Leader

• Summary: Developer Allen Bargi shared: Under AI-assisted programming, the focus shifts from "how to write code" to "clearly stating what is needed"—giving instructions, setting standards, and accepting results—which looks increasingly like team management.
• Source: allen.bargi.org · 2026-08-15
• Editor's Comment: One sentence hits home for many programmers. Technical ability in the AI era is no longer just typing speed, but the ability to define problems, break down tasks, and accept quality—skills that are transferable.

4. Has the AI Bubble Burst? The Community is Arguing

• Summary: Discussions around "whether AI valuations have peaked" continue to heat up, with the popular video Did the AI Bubble Just Pop? released. On one side, Anthropic's pre-IPO revenue is surging; on the other, concerns about high capital expenditures fuel bubble fears. Bull-bear divergence has reached its year-to-date high.
• Source: YouTube · 2026-08-15
• Editor's Comment: Bubble debates never miss, but there's a new variable this time: Giants' capital expenditures are real money. If returns don't meet expectations, cuts will happen faster than imagined. Watch the show, but don't stand too close.

5. SUSE AI 1.0 Released: Enterprise-Grade AI Open Source Suite

• Summary: SUSE released SUSE AI 1.0, a set of multi-layered enterprise AI solutions covering model serving, application frameworks, and deployment management, focusing on open source, controllability, and privatization. It targets enterprise customers unwilling to hand over data to cloud vendors.
• Source: SUSE · 2026-08-15
• Editor's Comment: The Linux veteran enters the enterprise AI platform market with the old playbook—open source + self-hosting + compliance. European companies have few choices in AI; SUSE aims to occupy this ecological niche.

6. Muxel: Opening Multiple Windows for AI Coding Agents

• Summary: Muxel is a terminal multiplexer for AI coding agents: supporting splitting multiple agent sessions across screens, pages, and drag-and-dock windows, managing multiple AI assistants like multiple terminal windows for parallel work.
• Source: Muxel · 2026-08-15
• Editor's Comment: One agent isn't enough? Then open a dozen. The toolchain is quickly filling the productivity gap for "multi-agent collaboration." The next battle is orchestration and preventing cross-talk.

III. Humanoid Robots

1. Comma.ai Releases PCIe Gen4 X4 to USB4 Dock, Firmware Open-Sourced

• Summary: Autonomous driving open-source player Comma.ai launched a PCIe Gen4 X4 to USB4 dock with fully open-source firmware. Developers can connect high-performance devices to vehicle computing systems, providing a more flexible hardware base for numerous open-source autonomous driving projects.
• Source: Phoronix · 2026-08-15
• Editor's Comment: Comma.ai has always followed the path of "handing the car computer to the community." The new dock lowers the barrier for self-developed hardware. What the robotics track lacks has never been models, but reliable edge computing carriers.

IV. Autonomous Driving

1. Real-World Comparison: Tesla FSD vs Rivian Autonomy+, Which is More Stable?

• Summary: CNBC reporters test-drove vehicles equipped with Tesla FSD and Rivian Autonomy+, horizontally comparing their "hands-free driving" experiences