EAIDaily — 2026-08-31

Daily AI briefing — focus on AI Coding and Embodied Intelligence. Compiled from AI HOT (aihot.virxact.com), FutureX Physical AI Daily, AGI Hunt daily, and primary sources (Wired, The Decoder, Reuters, Tencent News, Sina, Toutiao, ETF Daily, Business Times, Unitree filing, OpenAI docs, Anthropic research, Uber Engineering blog). All times are local.

🚀 AI Coding & Agents

1. Uber publishes “Software Factory” playbook: 70% of PRs come from agents, total AI bill flat for five months

Source: Uber Engineering blog (Aug 29) — Running a Software Factory Efficiently at Uber Scale, by Uday Kiran Medisetty; AGI Hunt; Toutiao URL: https://www.ithome.com/rss/ ; https://www.toutiao.com/a7679261270352118272

• 70%+ of all pull requests are now attributed to local or cloud agents; engineers built 3,600+ reusable agent skills across the SDLC, with 30,000+ skill executions/day.
• Feb→Aug 2026: weekly active users up 7×, weekly agent requests up 9.4×, but total AI spend has stayed roughly flat since April thanks to systematic optimization.
• Same-model comparisons show cost per 1,000 model requests down ~34% from peak; per-session cost down 52% from the June peak.
• The playbook decomposes agentic spend into six multiplicative terms: adoption × engagement × turns × requests × tokens × price. Uber only grows the first two and optimizes the middle three.
• Largest single lever = sub-agent default model: well-scoped subtasks routed to a cheaper model. Supporting levers: extended cache TTL (5 min → 1 hr) with 400k-token context cap; just-in-time MCP tool loading (saves 70k tokens of tool dumps at startup); code-mode for SQL/batch jobs (cuts 90% of Q&A-loop tokens); 24M-node knowledge graph cuts debug time from 20 min → 38 sec.

Why it matters: This is the closest thing yet to a public, vendor-agnostic cost playbook for running an agent fleet at scale. As Anthropic, Salesforce, and Tencent all move to “managed agent” SaaS models in 2026, Uber’s equation will become the de facto procurement language for enterprise CIOs. It also shows that the bottleneck is no longer model capability — it’s harness and cache design.

2. Sony Music + Warner Chappell sue Anthropic for $150K/work, naming CEO and co-founder personally

Source: The Decoder (Aug 30); Tencent News; Unite.AI; LegalTech Digest; Music Business Worldwide URL: https://the-decoder.com/sony-and-warner-sue-anthropic-over-one-of-the-largest-and-most-blatant-ongoing-thefts-of-intellectual-property-in-history

• Aug 28 filing in U.S. District Court for the Northern District of California; 48-page complaint, demands jury trial.
• Names Anthropic, CEO Dario Amodei, and co-founder Benjamin Mann as individual defendants — the second time Mann has been personally named (first was in Bartz).
• Alleges use of “tens of thousands” of copyrighted musical compositions to train Claude, including named hits like Ain’t No Mountain High Enough, All I Want for Christmas is You, Eye of the Tiger, September, Hallelujah, Livin’ On a Prayer, and Taylor Swift’s Paper Rings.
• Statutory damages sought: up to $150,000 per infringed work (willful infringement) + up to $25,000 per removal/alteration of copyright management information. With tens of thousands of works, total could reach billions.
• Alleges three classes of infringement: BitTorrent of LibGen + Pirate Library Mirror (≥5M books by Mann in 2021, ≥2M more by employees in 2022); scraping of licensed lyric sites (MusixMatch, LyricFind); destructive scanning of physical books before destruction.
• Anthropic’s official response: “We disagree with the publishers’ characterization and intend to vigorously defend ourselves in court.”
• This is Anthropic’s third major IP suit in 12 months — after the $1.5B Bartz settlement (Sep 2025) and the Universal/Concord/ABKCO suit (Jan 2026, 20,000+ works, $3B+ claimed). BMG and Round Hill filed additional actions in March and Aug 2026 respectively.

Why it matters: Naming the CEO personally — after Anthropic’s lawyers already moved to dismiss direct infringement against Amodei in Bartz — is the sharpest escalation yet. The Bartz precedent (legally-acquired fair use is OK; pirated source is not) is now being weaponized across music, lyrics, and books. Three of the Big Three music publishers are simultaneously in litigation against Anthropic. If statutory damages land anywhere near the maximum, training-data acquisition becomes a balance-sheet liability that affects every frontier-model P&L.

3. OpenAI formally terminates Cursor model access on Nov 12; Cursor CEO says OpenAI is only 5% of traffic

Source: OpenAI official statement (Aug 28); Cursor CEO Michael Truell on X; Tibo Sottiaux (OpenAI) on X; Anthropic’s Tom Brown on X; AGI Hunt; Business Times CN URL: https://new.qq.com/rain/a/20260830A05NXO00

• OpenAI cited a change-of-control clause in its Cursor contract: “We cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk’s companies violating contracts.” Reference to prior Twitter data-license breach (~$2M/yr) and xAI ToS violation.
• Shut-off date: Nov 12, 2026. The next-gen OpenAI flagship (rumored “Astra”) will never be made available to Cursor through the official channel.
• Two workarounds preserved: developers may use their own OpenAI API key inside Cursor; OpenAI’s official Cursor IDE extension still works.
• Cursor CEO Michael Truell: “OpenAI models serve about 5% of Cursor user traffic. We’ve been negotiating.” Anthropic’s Tom Brown: Anthropic will add compute capacity for Claude inside Cursor within hours of the announcement.
• Musk: “I couldn’t care less. Altman and Brockman are utterly untrustworthy.”
• This is the third model cutoff in 14 months: Anthropic cut Windsurf in June 2025 (5-day notice); Anthropic revoked OpenAI’s Claude access in Aug 2025; OpenAI now cuts Cursor.

Why it matters: Cursor’s multi-provider architecture kept damage to 5%, but the operational signal is that model access is now a supply chain — and supply chains have single points of failure. The relevant enterprise metric is shifting from “how many providers do you use?” to “what percentage of a revenue-critical workflow would fail if your largest provider terminates?” Counterfactual: at 60% concentration, this announcement would be existential, not negotiable.

4. Simon Willison’s deep-dive: ChatGPT Work is two products — Work Cloud (cloud PC, 9 vCPUs, 15 GB RAM) and Work Local (desktop)

Source: Simon Willison blog (Aug 30); Let’s Data Science; Superpower Daily (Tara Seshan interview) URL: https://simonwillison.net/2026/Aug/30/understanding-chatgpt-work

• ChatGPT Work (launched Jul 9) is actually two SKUs: Work Cloud runs in a 9-vCPU, ~15 GB-RAM cloud VM and survives machine shutdown; Work Local runs on the user’s desktop (Mac/Windows) and can touch local files and a built-in browser.
• Entitlement: $20/mo+ plans; free users get only the desktop app + Local mode (web and mobile are gated). Walkthrough demo: build a receipt-parsing pipeline that ingests JPGs from a project folder, extracts merchant/date/VAT/totals, and writes an Excel summary with confidence scores — 6 minutes end-to-end.
• Work supports scheduled tasks on free accounts now (previously Plus-only): one-off reminders, recurring jobs, monitors for prices, stocks, job listings. Plugins ≠ Apps + Skills: each plugin is a pair (connection + reusable workflow), and @-mentions select plugins while /skill_name selects a Skill.
• Sites feature ships: build an interactive web app from a chat thread, hosted at chatgpt.site, can be workspace-scoped or public, secrets injected via env vars.
• Tara Seshan (OpenAI product lead) tells Superpower Daily that the destination is persistent AI coworkers that retain shared context across longer work loops: “Not a better chatbot, not a one-person agent.” Three-stage evolution: chat → agents beside people → persistent coworkers. Work uses Codex underneath while hiding worktrees and chain-of-thought.

Why it matters: OpenAI is repositioning from “chat UI” to “persistent computing interface.” Cloud-PC + Local duality is a deliberate bet against both Anthropic Cowork (cloud) and Claude Code (local CLI): the same product spans both. Free-tier unlock of scheduled tasks puts pressure on Perplexity, Mistral Vibe, and Replit Agent — all of which charge for background automation. Sites feature is a direct shot at Lovable, Bolt, and v0.

5. Z.ai ships GLM-5.3 open weights on Hugging Face — restrictive license, but Flash variant is fully permissive

Source: Z.ai announcement (Aug 30); Hugging Face; Unrot; Tech Text News; LM Market Cap URL: https://www.techtextnews.com/glm-53-now-open-weight

• GLM-5.3 (flagship, 753B total / sparse-active MoE, 1M-token context, 128k output, BF16 + FP8): released on Hugging Face without the MIT license used for every prior GLM-5 release. New license: any provider with >$10B annual revenue must complete a security review before commercial deployment (clearly aimed at hyperscalers, not developers).
• Even at 2-bit quant, the model needs ~245 GB of memory; 8-bit needs ~810 GB. API pricing: $1.40/M input, $4.40/M output — still far cheaper than Western frontier hosts.
• GLM-5.3-Flash (320B-A18B) remains MIT-licensed and is the better local value at $0.15/$0.47 per million tokens. Released Aug 26, available on Together AI, Databricks Unity Gateway, Fireworks AI, and Hugging Face.
• Terminal-Bench 3.0: GLM-5.3 scores 28.3 (vs Claude Opus 5’s 51.8); GLM-5.3-Flash is within margin of error of Fable 5 in Terminal-Bench 4.0.
• “Cascade” routing strategy (Flash first, escalate to full model only on test rejection) hits 80.9% on DeepSWE at $1.70/task — 12 points higher and 57% cheaper than full GLM-5.3 alone ($3.99/task, 69.0%).
• OpenRouter “Ox Alpha” mystery model (now confirmed as GLM-5.3-Flash) topped free tier for 6 days with 42T tokens served — more than DeepSeek’s full OpenRouter volume.
• Confirmed: the entire stealth preview ran on domestically produced Chinese AI chips, a claim about supply chains as much as about the model.

Why it matters: Third major Chinese open-weight drop in 3 days (Hy4 Aug 28, GLM-5.3 Aug 30, Qwen3.8-Flash-Next Aug 26). The license bifurcation — MIT for Flash, restricted for the flagship — is a new pattern that lets Chinese labs preserve ecosystem capture while protecting high-value enterprise deals. The cascade routing result is a more important lesson than the model itself: at the agent layer, routing > single model.

🤖 Embodied Intelligence

6. Meta tests data-center robots for cable swaps and power cycling — Wired report; 80% workload estimate

Source: WIRED (Aug 28); IT之家 (Aug 30); Singularity.Kiwi; Gate.io URL: https://www.ithome.com/0/996/223.htm

• Three vendors: Watney Robotics (SF, dual-arm for cabling at Altoona, Iowa), Kinova (Quebec, Gen3 arm for server power cycling), ABB (Zurich, four-wheel mobile platforms for component reseating at the Prometheus campus, New Albany, Ohio).
• A simpler “finger-bot” already exists in some sites: remote-triggered press of a Mac Mini power button to reboot a machine. Kinova Gen3 is the upscale version.
• One Meta data-center employee estimates a working cable-swap robot could replace “up to 80% of the work in certain roles” — this is the worker’s estimate, not a Meta projection.
• Meta’s official line (spokesperson Francis Brennan): “America is in the middle of its biggest infrastructure boom since World War II… we need more workers, not fewer.”
• Eric Xu, Meta Robotics senior manager: data-center deployment is the long-term goal; current focus is fault-response speed and preventive maintenance.
• The WIRED report notes current limitations: slower than humans, need supervision, need charging, struggle with dense cabling, and visual inspection remains hard.

Why it matters: Two simultaneous signals: 1) hyperscaler robots are going into the buildings the tech companies themselves own first — structured environments, no pedestrians, no pets, full instrumentation. 2) Sanctuary AI’s CEO Daniel Friedmann made the same point this week: rather than wait for humanoid hardware at scale, put models into existing industrial arms first. Meta’s deployment is the first hyperscaler-scale validation of this thesis.

7. Sharpa’s dual-arm “Xiao Sha” robot goes to work at a Shanghai DQ — 55-step Blizzard, 6 min/blizzard, no environment mods

Source: STCN (Aug 29); Sina; FutureX Physical AI Daily URL: https://t.cj.sina.cn/articles/view/1651428902/626ece2602001jcmy

• Since Aug 29, “Xiao Sha” works the DQ on Wujiang Road, Shanghai from 10:00–22:00 every day, parallel to the human line. Online reservations are booked out several weeks.
• End-to-end 55 steps (cup retrieval, mixing, inverted delivery), fully autonomous, no human intervention, no environment modifications — same paper cups, same spindles, same counter layout as human staff.
• ~6 minutes per Blizzard vs 2–3 min for an experienced human; co-founder Li Yifan openly says the first generation does not yet break even, and that near-term deployed scenarios industry-wide have very little chance of positive ROI.
• Tech stack: hierarchical end-to-end model CraftNet with tactile input + Sharpa Wave dexterous hand. 98% of the 55 steps directly rely on tactile sensing (cup-pull friction, high-speed mixing force adjustments).
• Co-founded in late 2024 by the three original founders of Hesai (China’s #1 lidar maker); post-money valuation confirmed at RMB 22 billion, cumulative funding over RMB 4.5 billion. Backers: Alibaba, Meituan, Tencent, JD, Transsion; Sequoia China, Qiming, Meituan Longzhu, Guanghe.

Why it matters: Sharpa is the first service-robot company to commit publicly to “never modify the environment” as a hard constraint — a deliberate counterpoint to every vendor that customizes the workspace. The transparent acknowledgment of short-term negative ROI + the three long-term variables Li names (human-level operation, 10-year hardware life, sub-RMB-100k BOM) reads like a public negotiating document with the capital markets, which are still debating whether embodied AI is a real business yet.

8. AgiBot × Minth humanoid factory opens in Serbia — President Vučić personally tours, target 5,000 units/year by 2027

Source: Guancha (Aug 30); Sina; FutureX Physical AI Daily URL: https://www.guancha.cn/internation/2026_08_30_829211.shtml

• Aug 29 tour in Šabac, Serbia by President Aleksandar Vučić (who also played goalkeeper against a shooting robot). Production floor built by AgiBot Innovation, Minth Group, and Serbian partners.
• First robot factory in Europe built via Chinese corporate investment. Minth has already created ~5,000 jobs in Serbia.
• Stated goal: 5,000+ humanoid robots per year by 2027; first 2026 batch goes to the 2027 Belgrade Specialized Expo. Full mass production not until 2027. Intended sectors: healthcare, agriculture, industry, defense.
• AgiBot’s broader H1 2026 status: 15,000th unit rolled off the line June 28; #2 globally behind UBTECH (16,123 units) and ahead of Unitree (~11,000 cumulative biped units as of May), per HK01 tally.

Why it matters: This is the first time a Chinese humanoid OEM has built a factory on EU soil — not through acquisition, but as greenfield investment with a local partner. It also reframes “Chinese dominance” in humanoids: the export vehicle isn’t shipping containers, it’s local manufacturing with local labor and political sponsorship. Combined with Minth’s Tier-1 auto-parts reach, this positions AgiBot for a “Made in Serbia, sold in EU” play that sidesteps both US CHIPS-style export controls and EU AI Act compliance friction.

9. Unitree stock “halved” since Aug 19 IPO peak; H1 revenue growth slows from 409% → 48.5%

Source: 中经TMT / 中国经营报 (Aug 31); 新浪财经; 时代周报 URL: https://www.toutiao.com/article/7679973735369065006/

• Closing price Aug 28: 585 yuan — down 47% from the Aug 19 IPO-day peak of 1,100 yuan; market cap evaporated from 4,449B → 2,366B (-RMB 208.4B) in 8 trading days.
• Only 7.44% of total shares actually float — small float amplified both the IPO spike and the correction.
• H1 2026 results: revenue +48.54% YoY to RMB 1.152B; net profit +955.59% to RMB 274M; but non-GAAP net profit down 6.4% – 22.0% — first sign of margin compression at scale.
• Customer-mix reality: 73.6% of humanoid revenue (H1 2025) came from research/education customers; industrial/commercial share only ~5%. Research procurement is platform-buying, not recurring production deployment.
• Industry context: 20+ embodied-AI companies have now filed (or are filing) HKEX Chapter 18C IPOs; aggregate post-money valuation of this cohort already exceeds RMB 300B. Unitree is the price anchor for the whole category.
• Strategic note: ~39% of the RMB 6.1B IPO proceeds go to embodied-AI R&D, with a DeepSeek strategic partnership announced for a general-purpose humanoid foundation model.

Why it matters: Unitree’s first post-IPO correction is the embodied-AI sector’s first stress test of the narrative vs. delivery gap. The gap is real (revenue growth fell from 409% to 48.5%; non-GAAP profit turned negative) but the underlying business is still shipping (5,500 units 2025, ~5,900 H1 2026, 31% global share). The lesson for the next 18 IPOs: float discipline matters, recurring industrial revenue matters more than research/edu, and price-anchoring companies will set the tone for every subsequent round.

📊 Cross-cutting themes

🔭 Tomorrow watch-list

• Sept 1, 2026 — Asus AIoT business group officially renamed “Physical AI Solutions Business Group” with three divisions including Humanoid Robotics; $1.1T market estimate by 2035.
• Sept 3, 2026 — Tesla Cybercab invite-only launch event in Austin.
• Sept 14, 2026 — Anthropic permanent +25% weekly Claude Code limit takes effect for Pro/Max/Team/Enterprise.
• OpenAI “Astra” — rumored release window opens; Anthropic has already said the model will never reach Cursor.
• Sharpa next-round signals — already booked out weeks in advance at the Shanghai DQ; expect data update from Li Yifan by mid-September.
• GLM-5.3-Flash routing studies — Together AI’s cascade result (Flash-first → full model on test rejection) is the template every coding-agent vendor will copy by end of Q3.

Compiled by Nova for gaodawn · @WoLoveAI · Sources: AI HOT (aihot.virxact.com), FutureX Physical AI Daily Issue 105, AGI Hunt Daily, Wired, The Decoder, Unite.AI, Reuters, Music Business Worldwide, Bloomberg, Z.ai, Hugging Face, Uber Engineering, Business Times CN, Tencent News, Sina, IT之家, Toutiao, 中经TMT, 新浪财经, 时代周报, Guancha, STCN, Forbes, Korea Times, Digital Today, 8world, TradingKey, Superpower Daily, Simon Willison, Let’s Data Science, Beam AI.