EastVoice · AI · 2026-08-23 · Issue #45

Editor’s Note: Beijing’s robot games opened with 2,056 humanoid robots across 666 teams, and one sprint bot hit 9.32 seconds for 100m, under the human world record. On the model front, DeepSeek switched weekends to off-peak pricing, a mystery “Ox Alpha” model on OpenRouter keeps pointing back to Zhipu, and the $3 trillion AI-infrastructure debt question went mainstream in Chinese media.

Commentary: This was the week WRC stopped being a sideshow. Real factory lines, real delivery promises, and a STAR Market debut opening up 630% changed the conversation. The demo era is over, now we watch the balance sheets.

---

I. AI Foundation Models

1. DeepSeek Flattens Its Weekend Pricing Curve

   • Summary: Starting Aug 23, DeepSeek charges the off-peak rate all day on Saturdays and Sundays, scrapping the peak/off-peak split it introduced only six days earlier. The change covers all API calls, no time-of-day distinction on weekends.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: Weekend discounts read as consumer-friendly, but they are also load-shaping. Idle GPU hours are the one thing a lab like DeepSeek can always sell cheaper.

2. GLM-5.3 Ships Quietly, and That Is the Interesting Part

   • Summary: Zhipu released GLM-5.3 on Aug 14 and opened its API on Aug 19 at GLM-5.2 prices. The upgrade keeps the 5.2 base and pushes post-training effort into coding, long-horizon agents, and cybersecurity.
   • Source: TMTPost · 2026-08-22
   • Editor’s Take: A flagship launch with no buzz says something real: post-training tweaks no longer move the market. Wait for the next base model, not the next tuning run.

3. OpenAI’s Codex Passes 20 Million Weekly Users

   • Summary: Codex lead Tibo confirmed the coding agent crossed 20 million active users this week, following CNBC’s earlier report that OpenAI’s agent products reached roughly the same weekly-active figure.
   • Source: TMTPost · 2026-08-22
   • Editor’s Take: Twenty million users makes Codex the default answer to “which coding agent.” For rivals the fight is no longer features, it is distribution.

4. Mystery Model “Ox Alpha” Hits OpenRouter, Clues Point to Zhipu

   • Summary: An anonymous model appeared on OpenRouter on Aug 20, free for one week. Developer Ben Davis ran it on the DeepSWE benchmark and it beat Claude Fable 5, while community benchmarks suggest a Zhipu GLM lineage.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: Leaking a flagship under a pseudonym is now a marketing tactic. Whoever Ox Alpha really is, the DeepSWE score is the message.

5. OpenAI Cuts GPT-5.6 Sol Prices by More Than 20%

   • Summary: Over the next three months, GPT-5.6 Sol API and credit pricing drops more than 20%, a direct response to intensifying price competition.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: A 20% cut announced three months out is aimed at cheap Chinese API rivals as much as at anyone else.

6. Zhipu’s Tang Jie: Stop Counting Parameters, Start Counting Compute

   • Summary: In a long X post, Zhipu founder Tang Jie pushed back on “bigger parameters, better model,” arguing the real Scaling Law is about compute scaling.
   • Source: TMTPost · 2026-08-22
   • Editor’s Take: The man deciding where the next billion yuan goes says the field over-indexes on parameter counts. Follow the money, not the slide decks.

7. Doubao Input Method Lands on HarmonyOS

   • Summary: 字节跳动 (ByteDance)'s Doubao keyboard hit the App Gallery as a trial build, with voice input backed by Doubao’s speech model, supporting dialects and mixed Chinese-English input.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: A keyboard is a beachhead. The input method is where ByteDance’s models touch hundreds of millions of fingers every day.

8. Guiyu Technology: Robot Brains That Fly Into Danger First

   • Summary: QbitAI profiles 硅羽科技, a startup building embodied agents for hazardous work, sending robot “brains” into sites humans should not enter.
   • Source: QbitAI · 2026-08-22
   • Editor’s Take: Everybody races for the factory floor. Fewer go after jobs humans should not do at all, and that market has almost no competition yet.

9. Xinghaitao: From Model to Productivity

   • Summary: During WRC 2026, 星海图 (Xinghaitao) hosted a forum on embodied foundation models, whole-body evolution, and commercialization, gathering researchers, developers, and executives.
   • Source: QbitAI · 2026-08-22
   • Editor’s Take: The industry conversation has moved from “can it do the demo” to “who deploys it at scale.” Forums like this are the market forming in public.

10. The $3 Trillion Shadow Debt Behind the AI Buildout

    • Summary: TMTPost weighs the WSJ’s Aug 17 report on Alphabet, Amazon, Microsoft, Meta, Oracle, Nvidia, Broadcom and others, asking whether AI infrastructure debt, estimated near $3 trillion, is brewing a subprime-style crisis.
    • Source: TMTPost · 2026-08-22
    • Editor’s Take: The bill for the AI buildout is now on the table. Whether it is “subprime” or just “expensive capex” depends on revenue, and revenue has not caught up with the balance sheets.

II. AI Software & Applications

1. ByteDance Preps a Doubao Office Product Aimed at WorkBuddy

   • Summary: 字节跳动 (ByteDance) is expected to spin Doubao’s “work tasks” mode into a standalone app as soon as next week, directly targeting Tencent’s WorkBuddy. The Doubao web UI already shows a separate “new work task” entry.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: ByteDance has been slow to ship a serious office agent. This move says it plans to close that gap fast.

2. Apple Trims Siri and Vision Pro Teams by About 200

   • Summary: After cutting roughly 60 Vision Pro roles, Apple is also restructuring Siri, bringing total job cuts to about 200, per AppleInsider and Bloomberg. The Vision Pro gaming unit is largely being shut down.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: Two straight years of Siri being late to the AI party. The cuts are an admission, not a plan.

3. Anthropic Rumored to Raise $100B at a $2 Trillion Valuation

   • Summary: mitrade reports Anthropic could publicly file for an IPO as early as late August, with a raise of $100 billion and a valuation of $2 trillion, which would dwarf SpaceX’s record.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: A $2 trillion tag makes Anthropic the most expensive bet ever on “safety first” as a business model. Either the market is right, or this round is the top.

4. Man Detained in Sichuan for AI-Fabricated Harassment

   • Summary: Police in Sichuan detained a man who used AI-synthesized fake images to harass someone, per CCTV reports. One of the more visible enforcement cases under China’s deepfake rules.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: The rules have been on the books for a while. Now they are being enforced with actual arrests, and synthesis tools just got more accountable.

III. Humanoid Robots

1. UBTech Brings a Customer’s Real Factory Line to WRC

   • Summary: 优必选 (UBTech) recreated a client’s production line 1:1 inside the WRC exhibition and ran it live, arguing that embodied intelligence’s moat is deployment, not demo videos.
   • Source: QbitAI · 2026-08-22
   • Editor’s Take: The fastest way to tell who is real at a robot fair: check who dares to run an actual customer line in front of the audience. UBTech did.

2. Second World Humanoid Robot Games Open in Beijing: 2,056 Robots, 666 Teams

   • Summary: The games opened at the Ice Ribbon with 2,056 humanoid robots, up 138% in team count from the first edition, across 51 events. Beijing, CCTV, the World Robot Cooperation Organization, and the Asia-Pacific Robot World Cup council co-host.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: A sports competition is a strange but effective benchmark: it forces general-purpose hardware through events nobody scripted.

3. Galaxy General’s Tennis Robot Plays Human-Robot Mixed Doubles

   • Summary: 银河通用 (Galaxy General) showed off the 1.75m-tall robot 银河星仔, billed as the world’s first fully autonomous tennis-playing robot, running a LATENT planning algorithm trained with deep reinforcement learning instead of pre-programmed shots.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: A 1.75m robot returning real serves is the cleanest demonstration yet that reinforcement learning can replace scripting in physical sports.

4. Efort and Qizhi Take On the “Last Kilometer” of Factory Automation

   • Summary: Leiphone reports that 90% of factories still cannot adopt humanoid robots: oil, dust, temperature swings, inconsistent materials. 埃夫特 (Efort) and 启智 (Qizhi) are targeting exactly that gap, continuous and accountable work on real lines.
   • Source: Leiphone · 2026-08-22
   • Editor’s Take: The showcase floor is crowded, the dirty workshop floor is empty. The “last kilometer” is where the real market sits.

5. WRC 2026: The Game Moved From Brains to Balance Sheets

   • Summary: Leiphone’s dispatch from WRC argues the embodied-AI contest now hinges on unit economics, who can deliver at cost, more than on model capability.
   • Source: Leiphone · 2026-08-22
   • Editor’s Take: “Better brains” was last year’s pitch. This year it is “show the ledger.”

6. Unitree Scales Back Some Events to Focus on Volume Production

   • Summary: 宇树 (Unitree) said multiple of its robots compete at the games, some run by its own team, others by customers building on its platform, while it trims participation in some events to keep production capacity.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: Skipping events to keep production lines running is the clearest sign of where Unitree’s priorities sit: shipping beats medals.

7. A PMF Marathon: What This Year’s WRC Was Actually About

   • Summary: TMTPost’s dispatch notes the games opened on Qixi, with a robot proposal drawing crowds. The same day, 宇树 (Unitree) listed on the STAR Market, opening at ¥1,100, up almost 630% from its IPO price, with market cap briefly above ¥400 billion.
   • Source: TMTPost · 2026-08-22
   • Editor’s Take: A robot proposing on stage while its maker’s stock opens 630% higher, all on the same day. That is 2026 in one paragraph.

8. Unitree Robot Starts at Li Auto as a Tour Guide

   • Summary: A Unitree humanoid named 笨笨 (Benben) has been “hired” by 理想汽车 (Li Auto) as a guide and greeter at its facilities.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: Showroom jobs are low-stakes and high-visibility, exactly the kind of pilot that produces real deployment data.

9. “Can We Build Robots Is the Wrong Question, Dare We Sell Them to Consumers Is Right”

   • Summary: Shangwei’s Qiyuan Q1 research lead Peng Wenzheng argues the robot body is no longer the bottleneck, the pain is “strong body, weak intelligence” and thin content ecosystems for consumer robots.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: Consumer robots live or die on ecosystems, and ecosystems need a device people actually keep. Shangwei is betting that moment has arrived.

10. Rongyao Team’s Robot “Lightning” Runs 100m in 9.32 Seconds

    • Summary: In trials at the Beijing games, the Rongyao team’s robot Lightning covered 100m in 9.32 seconds with a peak speed of 14.5 m/s, beating the 9.58s human world record.
    • Source: ITHome · 2026-08-22
    • Editor’s Take: Faster than any human ever, in a trial, on a track. The headline is not the time, it is that the gap is closing in every measurable dimension.

IV. Autonomous Driving

1. Tesla Sets Sept 3 Cybercab Event in Austin, Public Test Rides Expected

   • Summary: Tesla confirmed a Cybercab launch event on Sept 3 in Austin, Texas, with first public test rides widely expected.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: After years of “robotaxi is coming,” Tesla has to show it can operate at scale in one city before anyone believes the roadmap.

2. Lynk & Co Recalls Vehicles to Replace Lidar, Free of Charge

   • Summary: 领克 (Lynk & Co) proactively recalled certain batches of the Lynk & Co 900 and other models after finding lidar-equipped cars could see degraded assisted-driving performance in specific scenarios. Owners get free lidar replacement under vehicle warranty.
   • Source: ITHome · 2026-08-22
   • Editor’s Take: A proactive recall on the sensor that drives the whole ADAS stack beats waiting for the first accident to name the pattern.

V. World Models / Physical AI

No entries this issue.

---

物界前沿 | EastVoice