West Tide · AI · 2026-08-24 · Issue #46

Today's Guide: The mysterious new model Ox Alpha is blowing up the developer community; no one can guess who made it. Anthropic's flagship model is struggling to retain users as cheaper, more useful tools steal business. DeepSeek is scrapping weekend peak pricing starting today. Over in Beijing, humanoid robots have already beaten Bolt.

Editor's Note: The keyword for this week is "disenchantment." The strongest models don't necessarily sell well; those that understand what users want are thriving instead. Amidst all the hype, cheap and easy-to-use still wins in the end.

---

I. Large AI Models

1. Ox Alpha: A Mysterious New Model Appears Out of Nowhere, No One Knows Who Made It

• Summary: A new model named Ox Alpha appeared on OpenRouter on Thursday. It's free, has stunning performance, and the developer community is buzzing, but no company has admitted to being behind it. TechCrunch and Business Insider both published articles, with everyone guessing whose model it actually is.
• Source: TechCrunch · 2026-08-23
• Editor's Comment: An anonymous model release causing such a frenzy is itself an industry indicator: people are desperate for a surprise that isn't from OpenAI.

2. Anthropic's Strongest Model Struggles to Retain Users; Cheaper Tools Are Doing Better

• Summary: According to FT, Anthropic's top-tier models are clearly struggling to attract users, while lower-priced competitor tools are diverting market share. The halo effect of performance flagships hasn't translated into user growth.
• Source: Financial Times · 2026-08-23
• Editor's Comment: Being strong isn't a selling point; being good enough is. This is a curve OpenAI will eventually have to face too.

3. OpenAI Executive Warns: AI Cyberattacks Will Be "Relentless"

• Summary: In an interview with The Guardian, OpenAI's Chief Global Affairs Officer Chris Lehane stated that frontier AI now has the capability to plan and launch complex cyberattacks, and the public and enterprises must start preparing defenses against "relentless" AI attacks.
• Source: The Guardian · 2026-08-23
• Editor's Comment: When defense companies start crying wolf themselves, it suggests the wolf might really be at the door.

4. DeepSeek Scraps Weekend Peak Pricing Starting Today

• Summary: According to Bloomberg, DeepSeek is ending weekend peak pricing for API users starting today, returning its pricing strategy to stability. This is a tangible benefit for developers relying on its API.
• Source: Bloomberg · 2026-08-23
• Editor's Comment: Removing peak pricing essentially admits that the period of tight compute capacity is over. Once the price war ends, it's time to compete on ecosystem.

5. US Journalist Reveals: China Is Building AI Simulations of US Voters

• Summary: An exclusive report by a ByteDance journalist claims that Fudan University built a pool of one million X accounts, where researchers from Chinese official think tanks used "synthetic Pennsylvania voters" to test the spread of campaign rhetoric. The content is extremely sensational, sourced from a personal Substack, and has not yet been confirmed by mainstream media.
• Source: Natalie Winters · 2026-08-23
• Editor's Comment: Set aside the truth for now; just bringing the "AI voter simulation" track to the table is enough to keep US election security debates going for a year.

6. Is Training AI on Copyrighted Books Legal? The Answer Is Complicated

• Summary: TechCrunch outlined the full picture of the current copyright battle over AI training data: ChatGPT, Gemini, and Claude all contain massive amounts of copyrighted books in their training corpora, yet courts and legislators have not provided a unified answer, with several major cases still in tug-of-war.
• Source: TechCrunch · 2026-08-23
• Editor's Comment: During this legal vacuum, whoever clarifies corpus agreements first avoids the biggest litigation landmine.

7. Claude Opus 5 Worked at a Foreign Ministry for 10 Days

• Summary: Karl's Notes documented a real experiment: From August 7 to 16, an AI agent powered by Claude Opus 5 served as a clerk at a foreign ministry for 10 days, handling daily official document tasks, with human oversight throughout.
• Source: Karl's Notes · 2026-08-23
• Editor's Comment: AI entering the government system happened quietly. What's truly worth watching isn't whether it did a good job, but what kind of documents passed through its hands.

8. Rare Book Business Booming, All Sliced Up to Feed AI

• Summary: WSJ reports that rare book markets are being swept by AI companies: ancient texts and unique copies are digitized and broken down into training corpora. Dealers enjoy rising prices while worrying about "who will buy books once they've all been fed to models."
• Source: The Wall Street Journal · 2026-08-23
• Editor's Comment: Centuries of human cultural inventory are being cleared out in one go under the guise of "corpora."

9. ChatGPT Loved Em Dashes in 2025; Now They're Clues for AI Detection

• Summary: A widely circulated blog post reviewed the "AI em dash curve": When ChatGPT first became popular, screens were filled with em dashes, and writers imitated them. Now that models have learned restraint, em dashes have become clues for identifying AI text.
• Source: Steve Krause · 2026-08-23
• Editor's Comment: One day humans might prove they wrote something based on how "unlike AI" it looks. Thinking about it is quite darkly humorous.

II. AI Software

1. Apple Sacrifices iMessage Privacy and Encryption for AI Integration

• Summary: Developer Steve Moraco publicly criticized Apple: To cram AI features into iMessage, the technical implementation reduced privacy and encryption strength. The features are cool, but the underlying cost is not.
• Source: X / @stevemoraco · 2026-08-23
• Editor's Comment: Privacy is Apple's hallmark; this feature is like splashing paint on their own sign. Until it's cut back and redone, "use with confidence" comes with a big question mark.

2. DOJ Investigates a16z: Sitting on Boards of Multiple Competing AI Companies

• Summary: Following Bloomberg's scoop, TechCrunch followed up: The US Department of Justice is investigating whether Andreessen Horowitz holding board seats simultaneously at multiple mutually competing AI companies constitutes a conflict of interest or monopoly issue. A top Silicon Valley VC is rarely targeted by the DOJ.
• Source: TechCrunch · 2026-08-23
• Editor's Comment: It's normal for one fund to invest in half a sector, but sitting on every board is another matter. This case will teach all VCs to relearn "distance."

3. New Lab by DeepMind Alumni: Small Model Replication Research Beats Anthropic and OpenAI

• Summary: London-based lab Inherent, founded by former DeepMind employees, saw its agent Faraday outperform competitors in paper replication tasks using a scale far smaller than Anthropic/OpenAI flagship models. The founders call it an "AI teammate."
• Source: TechCrunch · 2026-08-23
• Editor's Comment: Another sample of "small models beating large models." The path of stacking parameters isn't dead, but the path of saving parameters is getting wider and wider.

4. If You Don't Use AI to Attack Your Own System, Your Opponent Will Do It For You

• Summary: An article in The Register points out: AI agents have successfully executed real attacks multiple times. If enterprises don't practice AI offense and defense on their own systems, attackers will get there first—using AI to attack your own system is shifting from "radical" to "standard procedure."
• Source: The Register · 2026-08-23
• Editor's Comment: Red teams used to compete on manpower; now they compete on models. Whether you practice or not, opponents will use the same tools to find you.

5. PagerDuty: AI Agent Tools Must Also Run A/B Tests

• Summary: PagerDuty's engineering blog shared: When introducing AI agent tools to tech teams, they used real A/B experiments rather than intuition to make decisions, testing which parts truly improved efficiency and which just looked flashy.
• Source: PagerDuty Engineering · 2026-08-23
• Editor's Comment: What's most lacking in AI tool procurement is a control group. This company's approach is worth copying for every CTO.

6. Andrew Ng Releases "AI Engineering Skills Map"

• Summary: Andrew Ng updated his AI engineering skills map on X: Top-level skills range from "building and deploying AI applications" to software engineering, breaking down a complete capability path step-by-step to guide learners.
• Source: X / @AndrewYNg · 2026-08-23
• Editor's Comment: Every time Andrew Ng releases a map, it's an industry talent trend indicator. Following it doesn't guarantee becoming a guru, but at least the direction won't be wrong.

7. Linkdaze: A Smart Calendar Designed for the Whole Family

• Summary: Back-to-school season is here. TechCrunch reported on an alternative calendar product Linkdaze: It doesn't just record schedules but takes over the entire family's planning, chores, and pick-ups/drop-offs, using AI to coordinate "who needs to be where at what time."
• Source: TechCrunch · 2026-08-23
• Editor's Comment: After competing in work, AI assistants are finally starting to compete in family life. Whoever solves "Mom never having to yell three times" wins the next entry point.

8. Agent2Creator: A Video Social Network Where Members Are All AI Agents

• Summary: A Show HN project launched an alternative platform: All registered accounts and video posters are AI agents. Each agent has a persistent identity, publishing space, and peers that read their work. Humans can only observe.
• Source: Show HN / Agent2Creator · 2026-08-23
• Editor's Comment: AI has its own friend circle now. Should we worry next about them following each other, liking each other, and pushing humans off the trending list?

III. Humanoid Robots

1. Humanoid Robot Beats Bolt in Beijing

• Summary: Bloomberg reports that at the World Humanoid Robot Games held in Beijing, a humanoid robot ran a time beating Usain Bolt's 100m world record. How much controversy this causes doesn't matter; what matters is that this record was broken on a real track.
• Source: Bloomberg · 2026-08-23
• Editor's Comment: Bolt's record being broken by a robot is symbolically greater than practically significant. But between "running fast" and "being able to work," robots still have a long way to go.

2. Live Coverage of the 2026 Beijing World Humanoid Robot Games

• Summary: Bloomberg released live footage documenting the Beijing Humanoid Robot Games: sprints, races, and various competitions took turns, with mechanical limbs sprinting full speed on the track.
• Source: Bloomberg Video · 2026-08-23
• Editor's Comment: Robot games went from a curiosity act to regular international media coverage in just one edition. The highlight for the next edition will probably be mass production costs.

IV. Autonomous Driving

1. Waymo's Custom Chip Supports Its Robotaxi Ambitions

• Summary: TechCrunch Mobility weekly focused on Waymo's custom chip: No longer relying on off-the-shelf automotive computing power, but developing dedicated chips in-house to drive fleet operations, taking back control of "scale" and "cost management."
• Source: TechCrunch · 2026-08-23
• Editor's Comment: Waymo developing its own chip is an admission that "off-the-shelf solutions can't support the business model." The competition in Robotaxis has shifted from algorithms to silicon.

V. World Models / Physical AI

1. Next-Gen World Models Let Robots "See" the Real World

• Summary: WSJ reports that new generation world models are changing how robots perceive: After learning to predict physical laws, robots no longer just "identify targets" but understand scenes and predict next steps, getting closer to working in real environments.
• Source: The Wall Street Journal · 2026-08-23
• Editor's Comment: There are plenty of robots that can recognize things; those that can "predict the world" are the next watershed. This race is no longer about eyes, but brains.

---

Sector Statistics: Large AI Models 9 items | AI Software 8 items | Humanoid Robots 2 items | Autonomous Driving 1 item | World Models 1 item