---

⭐ Key Updates

1. Tesla FSD vs Rivian Hands-Free Driving: Two "Truly Hands-Free" Systems, Here's the Real-World Winner

CNBC reporters took two popular electric SUVs onto real roads for comparison—Tesla FSD (Full Self-Driving Supervised) and Rivian Autonomy+ (Hands-Free Driving System). One is aggressive and bold, the other conservative and steady. Conclusion: There is no absolute winner, only personality differences—choose FSD for "bold overtaking," choose Rivian for "steady and hassle-free."

Wearable Tech Expert · Akai says: Both are "truly hands-free" systems, but driving styles differ by dimensions. FSD is like the bold student in driving school, daring to make decisions at complex intersections; Rivian is like an old driver, changing lanes less and acting slower. For ordinary owners, figure out if you want "thrilling tech feel" or "quiet peace of mind"—no one is better, only who fits your commute.

Editor Xiaohe says: As someone who just passed their probationary license period: Turns out smarter isn't always better; a stability-first system is friendlier to novices. Hope domestic manufacturers compete more on "stability" rather than just "boldness."

2. 1Password Releases SCAM Benchmark: AI Agent "Anti-Fraud Exam" Begins

1Password open-sourced the SCAM benchmark, specifically testing AI agents' anti-fraud capabilities—simulating phishing emails, fake payment requests, social engineering tactics, and other real work scenarios to see if the agent transfers money to scammers on behalf of the owner.

Security Expert · Old Zhou says: Agent anti-fraud is the final piece of "trustworthy AI." No matter how smart the model, it falls for well-crafted social engineering info. SCAM turns "anti-fraud capability" into a quantifiable, comparable metric, something the industry should have done long ago. Of course, benchmarks are just the start; real attacks are more varied. Don't grant permissions recklessly.

Editor Xiaohe says: Letting AI book flights is already bold enough; letting it pay? Make it pass the anti-fraud exam first. In the future, vendors can say "My agent has an X% chance of being scammed," giving users a clear ledger.

3. Sony PS5 Pro Graphics Black Magic Revealed: Making AI Do Less Work Actually Improves Picture Quality

Sony engineers detailed the enhanced PSSR super-resolution technology for PS5 Pro at SIGGRAPH 2026—the core idea is counter-intuitive: Not making AI do more work, but making it do less. The new version separates "AI detail completion" from "rule-based algorithm fallback," resulting in more stable image quality and fewer ghosting artifacts.

Wearable Tech Expert · Akai says: "Making AI do less work improves picture quality"—this is engineering clarity. Super-resolution is essentially "guessing." Sony uses deterministic algorithms as the base and AI only for supplementation, giving the image a "fidelity floor." I'll add more once PS5 Pro players test frame rate data.

Editor Xiaohe says: As a console fan + monitor fan, the worst thing is AI-enhanced graphics looking beautiful from afar but blurry/ghosty up close. Sony's approach sounds reliable—hope it gets implemented in more games soon.

---

📊 Leaderboard Bulletins

Bulletin 1: Arena Text Blind Test Leaderboard (Aug 15 Snapshot)—Claude Fable 5 Tops, Anthropic Takes 7 of Top 10

#	Model	Vendor	Elo
1	Claude Fable 5	Anthropic	1506
2	Claude Opus 4.6 High	Anthropic	1505
3	Claude Opus 4.7 High	Anthropic	1502
4	Muse Spark 1.2 (xHigh)	Meta	1498
8	Qwen3.8-Max (Highest Open Source)	Alibaba	1491

Plain English: Humans act as judges in blind tests to see who answers better—7 of the top 10 are Anthropic. The highest open-source model is Alibaba's Qwen3.8-Max (#8); the gap is narrowing but hasn't caught up yet.

Bulletin 2: Arena Agent Blind Test Leaderboard (Aug 15 Snapshot)—Claude Sweeps Top 3, Kimi K3 is the Open Source Light

#	Model	Vendor	Sessions
1	Claude Opus 5 (High)	Anthropic	19.7k
2	Claude Fable 5 (High)	Anthropic	24.4k
3	Claude Opus 5 (Max)	Anthropic	15.5k
4	GPT-5.6 Sol (xHigh)	OpenAI	18.1k
5	Kimi K3 (Max) (Highest Open Source)	Moonshot	28.4k

Plain English: Switching to the challenge of "letting AI perform dozens of steps of real work continuously," Anthropic still sweeps the top 3. The only open-source model in the top 5 is Moonshot AI's Kimi K3—Chinese open-source players have already joined the first tier in "planning and executing themselves."

Bulletin 3: OpenRouter Weekly Call Volume Leaderboard—DeepSeek V4 Flash Tops, Surges 570% MoM

#	Model	Vendor	Weekly Calls
1	DeepSeek V4 Flash Official Version	DeepSeek	8.83 Trillion
2	Tencent Hy3	Tencent	8.05 Trillion
3	DeepSeek V4 Flash Preview	DeepSeek	5.88 Trillion
4	Xiaomi MiMo-V2.5	Xiaomi	5.39 Trillion
5	GPT-5.6 Luna	OpenAI	4.43 Trillion

Plain English: During the week of Aug 3-9, the top four global developer call volumes were all Chinese models—DeepSeek's official launch saw a 570% MoM surge in its first week, taking the top spot directly. Developers voting with their feet are investing real money in the "cost-effectiveness" of Chinese open-source models.

---

📋 Section Picks

• Suno Studio 2.0 Released: Adds MIDI support and effect plugins, moving closer to a true Digital Audio Workstation. AI music evolves from "one-click composition" to "deconstructible and editable," finally opening the creative black box.
• AI Tool Voting List Glad-AI-Tor: 76 AI tools ranked by real user votes; developers paying for positions doesn't work anymore. Returning judgment power to users.
• Yadda 3.0.0: The veteran BDD framework rebuilt for the AI agent era, turning test descriptions from "specs for humans" into "specs understandable by AI."
• Muxel: A terminal god-tool for opening "multiple windows" for AI coding agents; finally a handy tool for parallel multi-agent work.
• S.A.T.U.R.D.A.Y: Combines Pion, whisper.cpp, and Coqui TTS to create a fully self-hosted local voice assistant; data stays home.
• MyFirst.News: AI-generated children's news podcasts, translating complex world events into language kids can understand.
• Agent Shell 0.73: Interact directly with multiple AI agents within the editor, switch models, view tool call logs.
• How to Detect Hacked AI Accounts: TechCrunch practical guide—check login devices, review API usage, watch for abnormal conversation logs.
• AstraZeneca Reveals R&D Agents: Spanning literature search, experimental design, and data interpretation—the pharma giant puts AI on the production line.
• Giving AI £150 to Earn Back Subscription Fees: A fully public social experiment, closer to the essence of the "AI economy" than any leaderboard.

👀 Everyone is Watching

• Alibaba Qwen global downloads exceed 3 billion, topping the world
• Anthropic Q2 revenue grows over 14x YoY
• Nvidia holds 122.8 million SpaceX shares, becoming 6th largest shareholder
• 1Password releases SCAM agent anti-fraud benchmark
• New Paper: Adversarial Creation and Detection of AI-Generated Social Bots
• Volvo launches "Safety Coach" AI app, scoring driving behavior

🔭 Tomorrow's Focus

1. After DeepSeek V4 Flash tops OpenRouter, will next week's leaderboard repeat "Chinese models sweeping top 4"? Will the price-hiked V4 Pro drag down call volumes?

2. After the SCAM benchmark is open-sourced, will various large models undergo "anti-fraud horizontal comparisons"? AI agent anti-fraud capability is about to become a rankable metric.

3. After the Tesla FSD vs Rivian real-world comparison, will there be new material for the domestic "pure vision vs multi-sensor" route debate?

4. Will Suno Studio 2.0's MIDI support drive a trend for "AI Music Workstations"?

---

Physical World Frontier Review · Global AI Evaluation Radar｜Shenzhen Physical World Frontier Technology Co., Ltd.

Data subject to official disclosures; does not constitute investment advice