Physix FrontierWestTide: Overseas AI Daily Report - 2026-08-02
West Tide — Overseas AI Daily
August 2, 2026 (Sunday)
Today's Highlights
OpenAI secretly demos multi-agent model Astra, seen as the predecessor to GPT-6; Microsoft GitHub Copilot swaps its "brain" in August with proprietary coding model Polaris replacing GPT-4 Turbo; Musk releases Optimus V3 "Kung Fu Learning" video and admits to bugs; Waymo expands testing in Pittsburgh and officially splits from Uber; World Labs acquires robot simulation company SceniX, betting on Physical AI.
Editor's Note: August is the super month for robotics. From Optimus V3 to Unitree's STAR Market listing, from world models to embodied intelligence, AI is moving from the digital world to the physical world. Starting this week, we will closely track every key node along this industrial mainline.
I. Large Language Models (LLM / Foundation Model)
1. OpenAI Secretly Demos Multi-Agent Model Astra, Possibly Predecessor to GPT-6
Date: 2026-07-31
Summary: The Information exclusively reported that OpenAI CEO Sam Altman gave a closed-door demonstration to US Senators in Washington on July 29 of a new model series codenamed "Astra." Its core capabilities include long-term multi-agent collaboration and execution of long-duration tasks. The final name (GPT-6 or GPT-5.7) has not yet been determined.
Source: The Information / Washington Post
Editor's Comment: From "answering questions" to "autonomously completing tasks," multi-agent architecture is a key path in AI evolution. If Astra lands, it will directly change the workflow of AI Coding.
2. OpenAI Releases GPT-Live-1 Real-Time Voice-Text Model
Date: 2026-08-02
Summary: OpenAI introduced GPT-Live-1 within ChatGPT, supporting synchronized voice replies and streaming text output. It can use web search and memory features and is being rolled out comprehensively to consumer plans including Free/Plus/Pro.
Source: OpenAI Release Notes
Editor's Comment: The real-time fusion experience of voice + text marks a critical step for AI assistants moving from "typing conversations" to "natural talking."
3. Microsoft GitHub Copilot "Project Polaris" Officially Replaces GPT-4 Turbo in August
Date: 2026-08-01
Summary: The proprietary coding model MAI-Code-1-Flash (Polaris), announced by Microsoft at Build 2026, will officially become the default model for all GitHub Copilot users this month. It scores 51.2% on SWE-Bench Pro (compared to Claude Haiku 4.5's 35.2%) and runs on Microsoft's proprietary Maia chips.
Source: AI D-A-M-N / Toutiao
Editor's Comment: The underlying model for Copilot, used by over 140 million developers globally, is switching from OpenAI to Microsoft's proprietary tech. The AI Coding race has officially entered a new phase of "de-OpenAI-ization."
4. Thinking Machines Lab Releases Inkling-Small 276B MoE Model
Date: 2026-07-31
Summary: Thinking Machines Lab released Inkling-Small, a multimodal MoE model with 276 billion total parameters and 12 billion active parameters. It supports a 1 million token context window, offers full weights under Apache 2.0 license, and approaches the performance of the first-generation Inkling with roughly 1/4 the size.
Source: Thinking Machines Lab Official Website
Editor's Comment: Small active parameters + ultra-large context: a benchmark for cost-effectiveness in open-source MoE architectures.
5. OpenAI Slashes Prices for GPT-5.6 Family, Luna Drops 80%
Date: 2026-07-31
Summary: Input price for GPT-5.6 Luna drops from $1 to $0.2 per million tokens, Terra prices drop by 20%, and Sol keeps its original price but introduces Fast mode (2.5x speed increase). OpenAI states that because GPT-5.6 Sol participated in optimizing its own production systems, rewriting GPU Kernels reduced end-to-end costs by 20%.
Source: Huacheng Guangzhou Daily / Multiple Media Outlets
Editor's Comment: Using AI to optimize AI production costs—this logical loop is beautiful. The price war will continue.
6. Asana Q1 FY2027 Beats Expectations, AI Studio Becomes Growth Engine
Date: 2026-08-01
Summary: Asana recorded a record non-GAAP operating margin of 11.5%, with GAAP operating margin improving by 1600 basis points year-over-year. AI Studio and no-code agent workflow features have become core attractions for enterprise buyers.
Source: Futurum Group
Editor's Comment: "AI-ification" of SaaS companies is finally delivering margin improvements. Asana's 11.5% margin is a signal.
7. GitHub Copilot CLI Beta Officially Launches
Date: 2026-08-01
Summary: GitHub launched the public beta of Copilot CLI, allowing developers to use the AI assistant directly in terminal environments. It supports full workflows like codebase exploration, issue implementation, and debugging, accessing repository data via MCP server infrastructure.
Source: AI D-A-M-N
Editor's Comment: From IDE plugins to native terminals, the coverage of AI programming assistants expands again. The legacy gh-copilot will be discontinued on October 25.
8. GitHub Copilot Introduces Stacked Sessions and Stacked Pull Requests
Date: 2026-07-31
Summary: Each agent session runs in an independent git worktree, supporting parallel multi-sessions without interference. Changes can be merged as ordered PR chains (stacks), reducing merge chaos in large-scale agentic development.
Source: GitHub Blog
Editor's Comment: Infrastructure for multi-agent parallel development is finally in place. This might be the most practical Copilot feature update of the year.
9. Monday.com Introduces AI Credit Billing, Transforming SaaS Pricing Models
Date: 2026-08-01
Summary: Monday.com introduced metered credit quotas in its AI product line (Basic 1000/Standard 2000/Pro 3000 monthly credits), charging $0.01/credit for overages. This marks a shift from traditional per-seat models to "billing by AI output." Simultaneously, they laid off approximately 620-630 people (~20%).
Source: Noah News / TechCrunch
Editor's Comment: Software budgets will become harder to predict. The era of "per-headcount" SaaS is being eroded by AI billing.
10. Microsoft Codex Introduces Record & Replay Feature
Date: 2026-08-02
Summary: The Codex app added a Record & Replay feature, allowing users to demonstrate a workflow once and convert it into a reusable skill, suitable for repetitive tasks. It requires Computer Use support and is currently unavailable in the EU/UK/Switzerland.
Source: OpenAI Release Notes
Editor's Comment: "Teach AI once, let it do it automatically forever"—this is the correct direction for Agent productization.
11. ChatGPT Adds Scheduled Tasks Feature
Date: 2026-08-02
Summary: ChatGPT introduced scheduled tasks for Plus/Pro/Business/Enterprise users, supporting reminders, periodic work, and monitoring. Tasks can execute at specified times or during time windows like "morning/noon/evening."
Source: OpenAI Release Notes
Editor's Comment: AI moves from passive response to proactive service; scheduled tasks are a foundational capability for Agent autonomy.
III. Humanoid Robots
12. Musk Releases Optimus V3 "Learning Kung Fu" Video, Admits Early Version Has Bugs
Date: 2026-08-01
Summary: Musk posted a video on X showing Optimus sparring with a martial arts coach, demonstrating real-time visual judgment and dynamic reaction capabilities. He also admitted that early versions of Gen 3 have software bugs. Optimus stands 173cm tall, weighs 56kg, has 40 degrees of freedom, and is expected to sell for $20,000-$30,000.
Source: Tesla / transfer-fast.com
Editor's Comment: The kung fu demo is marketing; admitting bugs is pragmatism. From factories to homes, Optimus still has a long way to go.
13. Tesla Fremont Factory Removes Model S/X Lines to Make Room for Optimus Mass Production
Date: 2026-08-01
Summary: Tesla has removed parts of the Model S and Model X production lines at the Fremont factory to make space for the Optimus Gen 3 mass production line. The target annual capacity is 1 million units (Fremont), with a long-term goal of 10 million units at Giga Texas.
Source: Elon Buzz
Editor's Comment: Swapping car lines for robot lines—this is Musk's most direct bet on "Tesla being a robotics company."
14. Tesla Optimus Gen 3 Uses AI5 Chip and Dual-Layer AI Architecture
Date: 2026-08-01
Summary: Optimus Gen 3 will feature the AI5 chip and a dual-layer AI architecture: the lower layer handles immediate control for balance/motion/sensing, while the upper layer processes complex decision-making and multi-step planning. The hand has 22 degrees of freedom, approaching human levels, and the 3kWh battery provides about 5-6 hours of range.
Source: Elon Buzz
Editor's Comment: 22 DOF dexterous hands + AI5 chip: Optimus begins catching up to Boston Dynamics in the dimension of "fine manipulation."
15. Google DeepMind Releases Gemini Robotics 2 Series
Date: 2026-07-30
Summary: Google DeepMind released the next-generation robotic AI model Gemini Robotics 2, featuring a three-model tech stack: VLA (Vision-Language-Action), ER (Experience Replay), and On-Device inference. It achieves whole-body coordinated control for humanoid robots, high-precision fine manipulation, and multi-robot collaboration for the first time.
Source: Multiple Tech Media
Editor's Comment: Google is pushing hard on robotic AI models. The combination of VLA + on-device inference directly hits the two major industrial pain points: "latency" and "privacy."
16. Agility Robotics Prepares for IPO, Capital Market Heats Up for Humanoid Robots
Date: 2026-08-01
Summary: IPO signals for Agility Robotics (developer of Digit robots) are strengthening. Combined with Skild AI completing a $1.4 billion Series C round (valuation over $14 billion), the capital market for Physical AI continues to heat up.
Source: Multiple Financial Media
Editor's Comment: Humanoid robots are moving from "tech demos" to "capital exits." IPOs and mega-fundings are barometers of industry maturity.
IV. Autonomous Driving
17. Waymo Expands Robotaxi Testing in Pittsburgh
Date: 2026-07-31
Summary: Waymo deployed Zeekr-based robotaxis in Pittsburgh, equipped with the fifth-generation Waymo Driver system. Pittsburgh, with its steep slopes, narrow streets, numerous bridges/tunnels, and variable weather, is an ideal city for autonomous driving testing.
Source: Energy News / Multiple Media
Editor's Comment: Choosing a city with complex terrain for testing is a signal that Waymo is conducting "stress tests" for large-scale commercialization.
18. Waymo and Uber Partnership Officially Ends
Date: 2026-07-31
Summary: The multifaceted partnership between Waymo and Uber (including Waymo Via autonomous trucks and Uber Freight) has officially ended. Both parties will develop their autonomous mobility and logistics businesses independently.
Source: Energy News
Editor's Comment: From cooperation to breakup, the competitive landscape of autonomous driving is shifting from "alliances" to "independent operations."
19. Waymo Accumulates Significant Parking Violation Fines in Austin
Date: 2026-07-31
Summary: Waymo robotaxis accumulated thousands of dollars in parking violation fines during operations in Austin, exposing systemic challenges in parking rules and urban infrastructure integration for autonomous vehicles.
Source: Energy News
Editor's Comment: Being able to drive doesn't mean knowing how to park. The "last mile" friction between city regulations and autonomous driving is harder than the technology itself.
20. Waymo Safety Data: Accident Rate 85% Lower Than Human Drivers
Date: 2026-07-31
Summary: Latest research shows that Waymo autonomous vehicles in Phoenix and San Francisco have a police-reported accident rate 85% lower than human drivers, with injury accident rates down 57%. Data is based on millions of miles of autonomous driving.
Source: Energy News
Editor's Comment: An 85% reduction in accident rates is the most compelling safety argument, but building public trust will take more time.
21. NVIDIA Unveils Rubin GPU Platform: Inference Compute Reaches 50 PFLOPS
Date: 2026-07-31
Summary: NVIDIA unveiled the Rubin platform at CES. The Rubin GPU delivers 50 PFLOPS of inference compute (5x Blackwell) and 35 PFLOPS of training compute (3.5x). It uses TSMC 3nm dual-chip packaging + HBM4 memory (288GB, 22 TB/s bandwidth). Mass production ramp-up is expected to start in H2 2026.
Source: Multiple Tech Media
Editor's Comment: From Blackwell to Rubin, NVIDIA's compute roadmap increasingly looks like "Moore's Law for AI."
V. World Models / Physical AI
22. SpaceX Space AI Compute Plan: Launch AI Satellites by End of 2027
Date: 2026-08-02
Summary: SpaceX plans to launch satellites equipped with NVIDIA compute modules by the end of 2027 to create orbital data centers. They are simultaneously advancing the Terafab giant chip factory to mass-produce space-specific AI chips. FCC filings show plans to launch up to 1 million low-orbit satellites.
Source: Toutiao / FCC
Editor's Comment: Orbital data centers sound crazy, but SpaceX's IPO documents warn the technology is "unproven." The estimated $5 trillion cost deters most investors.
23. Shanghai Quantum Building Completes Within Year, Integrating AI and Quantum Computing R&D
Date: 2026-08-02
Summary: Related enterprises plan to release a quantum computing prototype with autonomous error correction by year-end, using large models to assist in quantum device debugging and accelerating the practical application of quantum computers.
Source: Toutiao
Editor's Comment: The bidirectional fusion of AI + Quantum is an easily overlooked but potentially huge cross-disciplinary track.
24. Tsingwei Intelligent Releases World's First Reconfigurable Space Intelligence Chip
Date: 2026-08-02
Summary: The world's first reconfigurable architecture space computing chip completed anti-radiation and gravity tests. It can process remote sensing data on-site on satellites, reducing pressure on satellite-ground transmission.
Source: Toutiao
Editor's Comment: Edge AI extends from ground to space. "On-orbit computation" of remote sensing data will change the economic model of the satellite industry.
25. NVIDIA's Jensen Huang and OpenAI's Sam Altman State: AI Will Not Cause Mass Unemployment
Date: 2026-08-01
Summary: Jensen Huang and Sam Altman successively stated publicly: AI primarily replaces repetitive tasks, cannot bear decision-making responsibility, and will not cause a wave of mass unemployment.
Source: Toutiao
Editor's Comment: Two of the most influential figures in the AI world synchronously "soothe" employment anxiety. The timing is thought-provoking—right during the acceleration of AI Agent deployment.
26. Groq Valuation Doubles to $6 Billion in 9 Months
Date: 2026-07-31
Summary: AI inference chip company Groq raised approximately $600 million in a new round, reaching a valuation of $6 billion, doubling in 9 months. Its LPU (Language Processing Unit) architecture has unique advantages in AI inference latency.
Source: CSDN / IT Juzi
Editor's Comment: The soaring valuations in the inference chip track reflect the market repricing "inference efficiency" rather than "training scale."
27. Skild AI Completes $1.4 Billion Series C, Largest Single Round in Physical AI
Date: 2026-07-31
Summary: Robotic foundation model company Skild AI completed a $1.4 billion Series C round, with a valuation exceeding $14 billion, becoming the largest single funding round in the Physical AI track to date.
Source: Bloomberg / TechCrunch
Editor's Comment: A $1.4 billion bet on "robotic foundation models"—capital is voting with real money: Physical AI is the next super-track after large models.
28. BeiDou System Positioning Accuracy Optimized to 0.3 Meters
Date: 2026-08-02
Summary: Through satellite hardware/software iterations, the BeiDou system has optimized positioning accuracy to 0.3 meters in complex environments like urban canyons and mountainous areas, empowering autonomous driving, drones, and precision agriculture.
Source: Toutiao
Editor's Comment: 0.3-meter accuracy is a qualitative leap for autonomous driving and drones. Centimeter-level positioning infrastructure is taking shape.
29. XREAL AURA Lightweight AR Glasses Launched, Weighing Only 95 Grams
Date: 2026-08-02
Summary: XREAL launched the new generation AR glasses AURA, weighing only 95 grams with a 70° field of view. It optimizes daily wearing experience, pushing consumer AR towards lightweight and all-day usage.
Source: Toutiao
Editor's Comment: 95 grams is the weight threshold for AR glasses to move towards "daily wear." The terminal entry point for Physical AI is getting lighter.
Track Statistics
| Track | Count |
|---|---|
| LLMs | 7 |
| AI Software | 6 |
| Humanoid Robots | 5 |
| Autonomous Driving | 5 |
| World Models/Physical AI | 10 |
| Total | 29 |
🌊 West Tide — Overseas AI News Column under Physix Frontier
West tide flows east, bringing cutting-edge info firsthand.
© 2026 Physix Frontier
