Xichao · AI · 2026-10-20 · Issue 103

Today's Briefing: The US is requiring federal systems to purge DeepSeek, adding another entry to the political ledger of models going overseas. Nscale and Anthropic are negotiating a $45 billion long-term cloud computing deal, while Anthropic simultaneously carves out 5% equity into a public interest fund. Chips, Agent security, and robotics each have one hard piece of news.

Editor's Note: What these piles of money will ultimately be repaid with — everyone's answer is different.

---

1. AI Large Models

1. US requires federal systems to purge DeepSeek

• Summary: The US issued a directive requiring federal agencies to remove DeepSeek from their systems and banning procurement of products and services from this Chinese company. The Department of Homeland Security separately notified federal civilian agencies to do the same.
• Source: SCMP · 2026-10-20
• Editor's Take: The directive governs data flows and procurement channels — it has nothing to do with whether the model is any good. For Chinese companies entering US public sector, the door is basically shut.

2. Nscale and Anthropic negotiating $45 billion cloud computing deal

• Summary: According to Bloomberg, UK data center developer Nscale is negotiating a cloud computing deal worth approximately $45 billion with Anthropic.
• Source: Bloomberg · 2026-10-19
• Editor's Take: Compute contracts keep getting bigger. The term length and default clauses say more about a deal than the total figure.

3. Anthropic puts 5% equity into public interest fund

• Summary: Anthropic announced it is putting 5% of the company's equity into a fund oriented toward public interest. Based on the $965 billion valuation from May this year, that portion is worth roughly $48 billion, with one use being to address job displacement caused by AI.
• Source: Guardian · 2026-10-19
• Editor's Take: The fund is registered as a wholly-owned subsidiary; the shares aren't donated. This money looks more like a political insurance policy.

4. Anthropic is preparing a model that better understands user intent

• Summary: The Information reports that Anthropic is internally developing a new model that can better judge users' true intent.
• Source: The Information · 2026-10-19
• Editor's Take: Intent understanding is a prerequisite for Agent deployment, and it's where vendors will differentiate themselves in the next round.

5. DeepSeek's V4.1-Flash update shakes up the industry

• Summary: The Information reports that the upgraded version of DeepSeek V4.1-Flash adjusted both inference speed and pricing, prompting the industry to re-evaluate the capability baseline of open-source models.
• Source: The Information · 2026-10-19
• Editor's Take: DeepSeek has walked the low-price, high-frequency-update path for a year, and followers still haven't found a way to crack it.

6. DeepSeek is changing the direction of Huawei AI chips

• Summary: The Information reports that DeepSeek's model adaptation needs are influencing Huawei's AI chip design direction, posing a new challenge to Nvidia.
• Source: The Information · 2026-10-19
• Editor's Take: Model vendors defining chip specs in reverse — this order has already taken shape in the domestic compute chain.

7. Four Gemini 4 agents beat Minecraft in 27 hours

• Summary: Google researchers published a paper in which four Gemini 4 Argon agents collaborated with division of labor and no human intervention, defeating the Ender Dragon in 27 hours.
• Source: Engadget · 2026-10-19
• Editor's Take: The agents' division of labor and mutual alerts are more interesting than the completion itself.

8. Meta's Prometheus cluster reaches gigawatt scale

• Summary: Meta's engineering team disclosed details of the Prometheus cluster, which has entered gigawatt scale, used to support training of the Muse series foundation models.
• Source: Meta Engineering · 2026-10-19
• Editor's Take: Multimodal training pushes power demand to the grid level; data center siting is now following substations.

9. AI proved the quasi-Riemann hypothesis, then what

• Summary: Quanta magazine reviewed the 377 results OpenAI announced on October 6, one of which handed the proof of the quasi-Riemann hypothesis to AI.
• Source: Quanta Magazine · 2026-10-19
• Editor's Take: The new problem mathematicians now face is how to verify a machine's proof.

---

2. AI Software

1. Tenstorrent raises $1.5 billion, valued at $6.2 billion

• Summary: Chip company Tenstorrent completed a $1.5 billion funding round at a post-money valuation of $6.2 billion. This round comes five years after Jim Keller became CEO.
• Source: TechCrunch · 2026-10-19
• Editor's Take: Inference is the keyword for this round of chip funding; the market beyond training is being priced separately.

2. AWS launches TorchNeuron, going head-to-head with Nvidia

• Summary: Amazon AWS released the TorchNeuron tool, making it easier for AI companies to migrate applications to its in-house chip Trainium.
• Source: TechCrunch · 2026-10-19
• Editor's Take: Migration cost has always been the biggest obstacle for in-house chips; the toolchain determines whether customers stay more than peak compute does.

3. Cloudflare cuts inference bills by 68%

• Summary: Cloudflare disclosed inference optimization details, claiming a series of engineering improvements reduced inference costs by 68%.
• Source: Cloudflare Blog · 2026-10-19
• Editor's Take: The same model deployed differently can save two-thirds of the cost. Most teams probably haven't seriously done this math yet.

4. Disney gets Character.AI, sidestepping its lawsuits

• Summary: Character.AI filed for bankruptcy this week; Disney acquired the team through asset purchase and taking on most employees, with litigation remaining with the original company.
• Source: Ars Technica · 2026-10-19
• Editor's Take: The bankruptcy process became a firewall for this deal; liability between AI content and copyright stays at the original address.

5. Hackers breach AI coding tool OpenCode, running up users' bills

• Summary: Open-source AI coding agent OpenCode admitted it was attacked, with users' linked AI accounts racking up high charges.
• Source: TechCrunch · 2026-10-19
• Editor's Take: What the Agent gets is a keychain, valid far beyond a single session. The broader the permissions, the broader the risk.

6. One hacker used AI agents to breach over 100 websites

• Summary: According to Reuters, a hacker used swarms of AI agents operating autonomously to breach over 100 websites, with John Deere and SharkNinja on the list.
• Source: The Verge · 2026-10-19
• Editor's Take: The cost was a few thousand dollars. The cost curves on both attack and defense sides are running in the same direction.

7. Eric Schmidt's 1% Anthropic stake is about to cash out

• Summary: Bloomberg reports that former Google CEO Eric Schmidt holds about 1% of Anthropic, and as the company's valuation climbs, the paper gains on this position are considerable.
• Source: Bloomberg · 2026-10-19
• Editor's Take: Early individual shareholders are sitting more comfortably in this round of AI valuations than institutions.

8. Nvidia releases new open-source models, purpose-built for always-on Agents on enterprise and PC

• Summary: Nvidia launched open-weight models for always-on Agents and edge computing, including the 120-billion-parameter Nemotron 3.5 Super VL.
• Source: TechCrunch · 2026-10-19
• Editor's Take: Moving Agents from cloud to device sells the whole hardware stack. Open-source models are the entry gift for this line.

9. World gives AI Agents identities, Zoom uses it against deepfakes

• Summary: Sam Altman's World project launched identity verification for AI Agents, and Zoom uses it to determine whether the other end of a call is a real person.
• Source: TechRadar · 2026-10-19
• Editor's Take: Once Agents have identities, accountability and auditing can land on specific entities. That's exactly what regulators want.

10. Apple tightens review, bans face-swap apps disguised as ordinary apps

• Summary: Apple updated its developer guidelines, banning apps that use AI to generate offensive content, and banning such apps from disguising themselves as ordinary tools to pass review.
• Source: TechCrunch · 2026-10-19
• Editor's Take: The same week, a Guardian investigation showed American teens can still install such apps. There's still a gap between rules and enforcement.

---

3. Humanoid Robots

1. JPMorgan deploys $5 billion in defense funding in first year, robot factories on the list

• Summary: JPMorgan completed $5 billion in venture investment in the first year of its US security and resilience program, covering robot factories, hypersonic missiles, and other directions.
• Source: CNBC · 2026-10-19
• Editor's Take: Manufacturing capacity is being recalculated as a defense asset. Robot production lines are getting capacity investment, unrelated to military orders.

2. Online exploration for robot Agents is being scaled up

• Summary: A new paper discusses scaling methods for online exploration of robot Agents, aiming to continuously collect experience in real environments without repeated retraining.
• Source: arXiv · 2026-10-15
• Editor's Take: Data collection efficiency determines the iteration pace of embodied models — more than body parameters do.

3. Using microscopic ground texture for localization, mobile robot precision takes another step up

• Summary: A new study uses microscopic ground texture to achieve high-precision mobile robot localization, providing an alternative in indoor scenarios where laser and vision solutions fail.
• Source: arXiv · 2026-10-15
• Editor's Take: Localization stability for warehouse and service robots often gets stuck on these unglamorous sensor approaches.

---

4. Autonomous Driving

1. Laplace Intelligence red-teams robots before they hit the road

• Summary: Laplace Intelligence specializes in adversarial testing of robots and autonomous driving systems, finding failure paths before mass production and deployment.
• Source: TechCrunch · 2026-10-19
• Editor's Take: Red-teaming is sinking from the model layer to the machine layer. The cost of one loss of control is far higher than one bad output.

2. DriveWorld3D uses evidence-constrained 3D world models for driving simulation

• Summary: A new paper proposes DriveWorld3D, combining 3D world modeling with generative simulation for autonomous driving training and validation in real road scenarios.
• Source: arXiv · 2026-10-15
• Editor's Take: The credibility of simulated scenarios determines whether they can replace road testing. Evidence constraints are a step in that direction.

3. Campus autonomous waste collection vehicle gets a two-tier system

• Summary: A study proposes the Mothership two-tier autonomous waste collection system, aimed at sidewalk-dominated campus environments, coordinating garbage collection vehicles and shuttle links.
• Source: arXiv · 2026-10-15
• Editor's Take: Low-speed, enclosed, fixed-route scenarios are where autonomous driving actually gets figured out first.

---

5. World Models / Physical AI

1. Trillions of operations per second is not a complete metric for physical AI

• Summary: AMD published a piece discussing how to evaluate compute for physical AI, arguing that treating TOPS as the sole metric misleads selection for real-world tasks.
• Source: AMD Newsroom · 2026-10-19
• Editor's Take: Moving from cloud to the physical world, latency and energy efficiency determine whether a system is usable more than peak compute does.

2. Does GPT-6 Astra actually have embodied intelligence

• Summary: A paper directly asks whether GPT-6 Astra already has embodied intelligence, examining item by item its performance boundaries on spatial reasoning and physical common sense.
• Source: arXiv · 2026-10-15
• Editor's Take: The value of such papers is splitting "looks like it can" and "actually can" into two separate questions.

---

Track stats: Large models 9 · Software 10 · Humanoid robots 3 · Autonomous driving 3 · World models/Physical AI 2, total 27.