West Tide · 2026-08-10

Today's Intro: OpenAI's largest pre-training model "Doug" exposed, base scaling restarts. Tesla permanently discontinues Model S/X, Fremont factory converted for Optimus Gen 3 production line. Amazon Zoox steering-wheel-less Robotaxi set for paid operations in Las Vegas soon.

Editor's Note: Mass production signals for the humanoid robot track are densely packed—Tesla factory conversion, Unitree IPO subscription, Zoox commercial operations. Three clues jointly point to one judgment: H2 2026 is the critical window for physical AI moving from Demo to Delivery.

I. Large AI Models

1. OpenAI's Largest Pre-training Model "Doug" Exposed

• Date: 2026-08-09
• Summary: X user ChrisGPT leaked that OpenAI is advancing a new round of large-scale pre-training codenamed Doug. SemiAnalysis's July memo confirmed "OpenAI has overcome pre-training issues." Doug is not the same model as GPT-6 and is expected to launch by November at the latest.
• Source: Jiqizhixin, SemiAnalysis, Ifeng Tech
• Editor's Comment: For the past two years, OpenAI drove capability growth via RL and post-training. Doug signifies a restart of scaling for the base model itself. If true, this will be the first true generational leap in base models since GPT-4o.

2. OpenAI Retires Atlas Browser—Capabilities Merged into ChatGPT and Codex

• Date: 2026-08-09
• Summary: OpenAI integrated Atlas's browser functions into the ChatGPT desktop app and Codex. Atlas ceased functioning on August 9, requiring users to manually migrate data.
• Source: OpenAI Release Notes
• Editor's Comment: The standalone Atlas browser product survived only a few months before being recalled, indicating that the AI-native browser path is difficult; embedding it back into main products is more practical.

II. AI Software

3. Apple Limits Bug Bounty Submissions—Too Many AI-Generated Vulnerability Reports

• Date: 2026-08-08
• Summary: Due to an excessive number of vulnerability reports generated by AI "hallucinations," Apple limited external submission quantities. An Italian startup discovered over 50 macOS vulnerabilities in 3 weeks but couldn't submit them all due to limits.
• Source: Multiple media outlets
• Editor's Comment: AI's speed in discovering vulnerabilities has exceeded human review speed. This is another typical case of the "double-edged sword" effect in the AI security domain.

4. Anthropic Claude Mythos Vulnerability Discovery Capability Draws Attention

• Date: 2026-08-08
• Summary: Anthropic's Claude Mythos model can discover critical security vulnerabilities, with Apple being one of its partners. Since the model has both the ability to find vulnerabilities and the potential to exploit them, it is currently limited to select enterprise use.
• Source: MacRumors, multiple security media outlets
• Editor's Comment: The "capability-risk" paradox of AI security researchers—an AI capable of finding vulnerabilities is equally capable of exploiting them.

III. Humanoid Robots

5. Tesla Converts Fremont Factory for Optimus Gen 3 Production Line—Model S/X Permanently Discontinued

• Date: 2026-08-09
• Summary: Tesla permanently removed Model S and Model X production lines from the Fremont factory, replacing them with the first dedicated Optimus humanoid robot production line. Gen 3 plans to start initial production in Q3, focusing on validating manufacturing equipment and component reliability.
• Source: Elon Buzz, Tesla Q2 2026 Earnings
• Editor's Comment: Discontinuing two classic car models to make way for robots—this is the most substantive signal of Tesla's transformation from a car company to a robotics company.

6. Musk Releases Terafab Render Video—Optimus + Robovan First Joint Appearance

• Date: 2026-08-09
• Summary: Musk released a new Terafab video on X, featuring Optimus robots and Robovan autonomous vehicles together for the first time. Terafab is planned for Texas, approximately 4 km long, with a current investment of $16.8 billion. Optimus targets large-scale mass production by summer 2027, with a final annual capacity of 10 million units.
• Source: IT Home, Musk's X Platform
• Editor's Comment: $16.8 billion investment + 10 million unit annual capacity target—Musk's bet on the robotics track is getting bigger.

7. Tesla Optimus Price Target $20k-$30k

• Date: 2026-08-09
• Summary: Musk stated that Optimus prices will drop to $20k-$30k after large-scale mass production, with some components shareable with Tesla vehicles (batteries, electronics, motors, etc.).
• Source: 36Kr Overseas, multiple media outlets
• Editor's Comment: $30k ≈ 200k RMB—if achievable, the price threshold for humanoid robots will be lower than a mid-range sedan.

IV. Autonomous Driving

8. Zoox Las Vegas Commercial Robotaxi Starts Paid Passengers Next Week

• Date: 2026-08-08
• Summary: Following NHTSA approval for commercial deployment of steering-wheel-less Robotaxis, Amazon-owned Zoox confirmed it will begin officially charging the public for driverless taxi services in Las Vegas next week. Vehicles have no steering wheels or pedals, feature a four-seat face-to-face layout, and drive bidirectionally.
• Source: Edinburgh News.Net
• Editor's Comment: Only a week from approval to actual operation; Zoox's execution is noteworthy. ⚠️ Single source, pending further verification.

9. Tesla Cybercab "Unboxed" Manufacturing Strategy Exposed

• Date: 2026-08-09
• Summary: Tesla's Q2 earnings revealed that Cybercab will adopt the "Unboxed" manufacturing strategy—splitting the vehicle into multiple modules for parallel production and assembly, significantly shortening production cycles. Giga Texas is preparing the Cybercab production line.
• Source: Elon Buzz
• Editor's Comment: If successful, the Unboxed strategy will redefine the efficiency ceiling of automotive manufacturing—but the engineering challenges are equally huge.

V. World Models / Physical AI

10. Tesla Digital Optimus Concept—Virtual Simulation Accelerates Robot AI Training

• Date: 2026-08-09
• Summary: Tesla proposed the Digital Optimus concept, training and testing neural networks in virtual environments. Software iterations expand rapidly in simulation, while physical robots focus on validating mechanical reliability and real-world performance.
• Source: Elon Buzz
• Editor's Comment: The dual-track mode of simulation training + real-machine validation is industry consensus. Tesla's advantage lies in the simulation infrastructure accumulated from FSD.

11. Tesla Optimus Training Data Loop—Optimus Academy Internal Training System

• Date: 2026-08-09
• Summary: Tesla established the Optimus Academy internal training system: Human demonstration → Teleoperation imitation → Factory deployment → Data recovery → Neural network update → OTA push. Every deployed robot serves as a training data source.
• Source: Elon Buzz
• Editor's Comment: The FSD data flywheel logic is being replicated in the robotics domain—this is Tesla's core competitive moat.

QC pass rate 42.9% (27/63). Lower weekend output this issue is normal. All above content passed two-stage quality control review.