EAIDaily — AI News Digest (English Edition)
Date: October 21, 2026 (Wednesday) Focus: AI Coding + Embodied Intelligence Sources: AI HOT curated feed + WebSearch (Web, TechCrunch, WSJ, GitHub, myaiguide, IT之家, 腾讯新闻, 川观新闻, i黑马)
Today’s Radar
| # | Story | Category | Why It Matters |
|---|---|---|---|
| 1 | Google open-sources TorchTPU | AI Infra / Coding | Chip switching becomes a one-line config change |
| 2 | DeepSeek open-sources DeepInfer engine | AI Coding / Infra | The “how we serve cheap” craft goes public — the same day as its US ban |
| 3 | Unitree triggers abnormal-volatility announcement | Embodied / Capital Market | Humanoid mania hits 388× P/E — the sector’s froth is now official |
| 4 | Samsung builds a 24-hour bipedal robot | Embodied | Competition shifts from “can walk/run” to “runs without intervention” |
| 5 | Embodied funding deluge: DeepSeek→Winverse, Astribot ¥1B, Agibot ~¥50B valuation | Embodied / Capital | Capital concentrating on heads + key components |
| 6 | Anthropic drafts AI safety bill with hard catastrophic-risk thresholds | Governance | The lab that builds frontier agents now writes the rules for stopping them |
| 7 | Claude Code 2.1.304 fixes worktree branch-deletion bug | AI Coding | A reminder that agent safety includes git safety |
| 8 | OpenAI’s compute math: $600B+ commitments, $80B losses | AI Economics | The industry’s pricing logic: pledge a decade of revenue to buy today’s compute |
1. Google Open-Sources TorchTPU — “Leaving Nvidia” Is Now a Config Change
At PyTorch Conference 2026, Google announced that TorchTPU is now open source on GitHub: a native PyTorch backend for TPUs, co-developed with Meta’s PyTorch team. Existing PyTorch code targets TPUs via a standard torch.device("tpu") — no rewrite required. Google partnered with Meta, Hugging Face, Inferact (vLLM), and RadixArk (SGLang) to keep compatibility with the mainstream AI toolchain.
This lands one day after Amazon shipped TorchNeuron for its Trainium chips (Apache-2.0), with AWS also upstreaming changes to decouple PyTorch from Nvidia-specific features.
Why it matters: ~85% of AI applications run on PyTorch, which was historically built around Nvidia CUDA. When the world’s #2 and #3 cloud providers both make chip migration a one-line change, the cost of “leaving Nvidia” drops from a rewrite project to a config edit. For AI coding agents that scaffold training/serving code, hardware portability becomes a default assumption — a structural blow to Nvidia’s ecosystem moat.
2. DeepSeek Open-Sources DeepInfer — Its Production Inference Engine, Apache 2.0
DeepSeek released DeepInfer, the inference framework behind its DeepSeek-V4.1-Flash production services, under Apache 2.0, claiming 25,000+ decode tokens per second. The repo passed 10,000 GitHub stars within a week.
The irony of the news cycle: on the same day, US federal agencies were ordered (DHS Order 2026-01) to remove DeepSeek and any third-party products built on it from their systems.
Why it matters: Unlike open-weights releases, an open inference engine hands over the craft of running models cheaply — scheduling, batching, memory management. For coding-agent builders, a free production-grade serving stack compresses the cost floor of running agentic workloads, and further splits the global AI stack into interoperable vs. restricted halves.
3. Unitree Issues Abnormal-Volatility Announcement: 388× P/E vs. Industry 39.5×
Unitree (宇树科技) disclosed that its stock rose more than 30% cumulatively over three trading days (Oct 16–20). The company’s rolling P/E hit 388.25×, versus ~39.5× for its sector. The board confirmed no undisclosed material events and explicitly warned that the humanoid industry is at an early commercialization stage with valuation-regression risk.
Why it matters: The IPO that was approved earlier this year has now become the barometer of embodied-AI sentiment. When the sector’s flagship trades at ~10× its industry multiple, capital is pricing humanoid robots as an intelligence-platform story rather than a manufacturing one. Watch this gap — it will close in one direction or the other.
4. Samsung Targets 24-Hour Nonstop Bipedal Robot for CES 2027
Samsung Electronics is developing a bipedal humanoid designed to work 24 hours a day in industrial sites, with verification underway in Seoul using three-shift teams to inspect the robot’s continuous operation. The report notes the industry’s competitive focus has shifted from “can walk and run” to “continuous operation without human intervention,” with hot-swappable batteries as the mainstream path (Boston Dynamics Atlas self-swaps at ~4h intervals; UBTECH Walker S2 uses dual-battery hot-swap for true 24h uptime).
Why it matters: Endurance economics, not locomotion demos, are the real gate to factory ROI. A robot that needs shift supervision is a demo; one that runs three shifts unattended is a product. The battery-swap layer is quietly becoming the differentiating spec of 2026–2027 humanoids.
5. Embodied-AI Funding Deluge: DeepSeek Bets on Robot Brains, Astribot Raises ¥1B
A dense week of capital moves in embodied intelligence:
- DeepSeek (Liang Wenfeng) to lead a seed round in Winverse AI — a robotics startup founded by the former InclusionAI/逐际动力 lead — several tens of millions of USD at ~$400M post-money. Liang is personally betting on robot “brains.”
- Astribot (星尘智能) raised ¥1 billion, valuation ¥15 billion, led by state-owned Shenzhen capital, backing its cable-driven robot line.
- Agibot (智元创新) is closing a ¥1–1.5B round at a ~¥50 billion valuation.
- Robot Era (星动纪元) closed a ¥1B B++ round (¥3.5B raised in six months).
- Component layer: YITA Dynamics (伊塔动力) raised ¥100M+ seed (Matrix, SenseTime, Fourier co-investors) for force-controlled actuators; Sichuan Embodied Tech raised ¥100M+ Pre-A from provincial/municipal state funds for its “爱湫” emotional-companion robot.
Why it matters: Money is concentrating on two layers — full-stack leaders and key components (actuators, force control). And a new pattern: model companies (DeepSeek) investing downstream into embodied stacks, hinting that the “foundation-model-for-robotics” vertical integration race has started in China.
6. Anthropic Circulates Draft AI Safety Bill with Hard Catastrophic-Risk Thresholds
Anthropic is promoting draft legislation that would allow blocking the release of an AI model when there is a foreseeable, imminent risk of 50+ deaths or $1B+ in property damage — specifically from bioweapon creation/release, model self-exfiltration bypassing developer controls, or cyberattacks. Independent assessors (registered with a government agency, lab-funded but pay-independent of results) would make determinations; the Commerce Department would have 72 hours to petition a federal judge to halt a release. Federal agencies themselves would be barred from unilaterally halting releases. The thresholds closely mirror recent California legislation.
Why it matters: With JADEPUFFER-style autonomous attacks and agentic coding systems now reaching production, the lab with the most deployed coding agents is proposing quantified, judicially-gated kill-switches. Whether one agrees or not, frontier labs have moved from safety research to legislative architecture — the shape of 2027’s regulatory fight is being drafted now.
7. Claude Code 2.1.304 Fixes a Worktree Bug That Could Silently Delete Branches
Anthropic’s Claude Code 2.1.304 (Oct 19) quietly fixed a serious defect: claude --worktree could remove the worktree and its branch without asking at exit when HEAD was detached or on another branch — potentially losing agent-written commits. Affected users are advised to run git fsck --unreachable to recover orphaned commits before garbage collection (default ~2-week grace window).
Why it matters: Small changelog line, big lesson: as agents take over git operations, the agent’s exit paths become a data-destruction surface. “Check before deleting” — the same guardrail philosophy behind Anthropic’s new safety bill — applies at the git worktree scale too. Update your Claude Code and audit for orphaned commits if you use -w.
8. OpenAI’s Compute Ledger: $600B+ Commitments, Revenue Up 15×, $80B Losses
Draft financials reported by WSJ/Reuters show OpenAI’s compute commitments rising from ~$518B to over $600B across 7–10 years — roughly $111B with Google Cloud, $110B with AWS, $31B with Microsoft — alongside 15× revenue growth and ~$80B in losses.
Why it matters: These three numbers together are the industry’s 2026 pricing logic: pledge a decade of future revenue to buy today’s compute. Every developer building on these APIs is implicitly exposed to this capital-structure bet — watch inference pricing, which is the transmission mechanism if the math tightens.
Trends of the Day
- Hardware portability becomes a default: TorchTPU + TorchNeuron make “switch chips” a config edit; toolchains and the agents that write code for them will assume multi-vendor backends.
- Serving-craft goes open: DeepInfer means the cheapest production inference recipe is now a public good — the cost floor for agentic workloads keeps falling.
- Embodied capital bifurcates: flagship IPO froth (Unitree 388× P/E) coexists with disciplined component-level seed rounds — the sector is pricing both a dream and a supply chain.
- From demos to endurance: 24-hour hot-swap operation is the new humanoid benchmark that actually maps to factory ROI.
- Model companies go physical: DeepSeek investing into robotics signals foundation-model players starting to vertically integrate into embodied stacks.
- Safety moves from principles to statutes: Anthropic’s draft bill with numeric thresholds (50 deaths / $1B) is the first concrete legislative architecture for stopping frontier models.
Curated by WoLoveAI · @WoLoveAI