Xichao · AI · 2026-11-05 · Issue 119

Today's Briefing: Uber plans to have thousands of driverless cars on its platform by 2027, while Waymo confirmed in the same week that it will end its exclusive partnership in Austin and Atlanta in January 2028. There was another number on Uber's earnings call: last quarter, trips completed by AI agents totaled 4. Anthropic turned "which model produced this output" into an open-source project called FlashProve, answering it with zero-knowledge proofs.

Editor's Note: The platform treats driverless cars as supply; driverless cars treat the platform as a channel. Same ledger, but the two sides don't get the same number.

---

1. Humanoid Robots

1. RoboScope Kit: A unified interface for embodied AI evaluation

• Summary: This arXiv paper comes from the robotics side, proposing RoboScope Kit, which folds the various homegrown scripts each lab writes for real-robot evaluation into a unified interface. The goal is to make evaluating embodied policies faster and horizontally scalable. The abstract emphasizes two things: speed and reusability.
• Source: arXiv · 2026-11-04
• Editor's Take: Without unified evaluation, demo videos aren't comparable. Tools like this aren't sexy, but they determine whose scores count later.

2. Open robot swarms use blockchain DAOs for their own access control

• Summary: The paper discusses how robots belonging to different companies can share space and tasks. The approach writes access permissions into a blockchain-based DAO, with the swarm voting and adjudicating on its own, without relying on a central scheduler.
• Source: arXiv · 2026-11-04
• Editor's Take: Once a swarm scales up, who can command which robot becomes a permissions problem. On-chain adjudication can settle ownership, at the cost of latency and a transaction fee.

3. CoReMatch: Coordinate refinement for real-time robot feature matching

• Summary: This paper tackles an old problem in robot vision: images and 3D space not lining up. CoReMatch uses a recurrent structure for continuous coordinate refinement, pushing matching accuracy into the real-time usable range.
• Source: arXiv · 2026-11-04
• Editor's Take: Among the causes of grasping failures, visual misalignment accounts for a far larger share than models being too small. Improvements at this layer rarely make headlines.

4. Robohub October roundup: Biohybrid robots push off the ground and jump

• Summary: Robohub's monthly roundup collects a batch of October robotics research and social topics, including a biohybrid robot that pushes off the ground after stimulation. The roundup also lists industry deployment and cultural discussions.
• Source: Robohub · 2026-11-04
• Editor's Take: The value of a monthly roundup is filtering out noise. To really see the direction, watch which type of project is still in it three months running.

---

2. Intelligent Driving

1. Uber wants thousands of driverless cars by 2027; Waymo is already at that scale

• Summary: InsideEVs reports that Uber aims to have thousands of robotaxis on its platform by 2027. The report uses Waymo as a comparison: the latter's existing fleet is already near that scale, and it operates them itself.
• Source: InsideEVs · 2026-11-05
• Editor's Take: Uber's numbers have to be assembled from partners; Waymo's numbers sit in its own garage. Same few thousand cars, completely different control.

2. Waymo confirms split with Uber in January 2028

• Summary: Waymo co-CEO Tekedra Mawakana told Bloomberg that starting January 2028, Waymo will end its exclusive partnership with Uber in Austin and Atlanta, and vehicles will no longer go only through Uber's entry point.
• Source: TechCrunch · 2026-11-04
• Editor's Take: Once the exclusivity ends, Waymo has to take orders itself. What it lacks isn't cars; it's a reason to pull passengers out of Uber's icon.

3. Tesla complained about in Korea: most cars can't use FSD at all, yet it's still sold as a selling point

• Summary: Electrek reports that a Korean consumer group has filed a complaint with regulators, alleging Tesla knew most vehicles on sale locally cannot legally enable FSD yet still pitched it as a selling point during sales, affecting tens of thousands of owners.
• Source: Electrek · 2026-11-04
• Editor's Take: Whether a feature works depends on local regulatory approval. Leaving approval risk to owners will eventually come back to the brand.

---

3. World Models · Physical AI

1. EdgeDAE: Compressing diffusion action experts to real-time on-device

• Summary: The paper accelerates action-generation models like diffusion action experts, aiming for real-time performance on edge hardware to serve physical AI control loops. The authors categorize it under hardware architecture.
• Source: arXiv · 2026-11-04
• Editor's Take: The more diffusion-like an action model is, the more expensive inference gets. Robots can't wait for those 200 milliseconds in the cloud.

2. Compression Forcing: Compress history into hierarchical tokens so world models can stretch longer

• Summary: This paper proposes Compression Forcing, which slices long history into hierarchical tokens, letting world models stay controllable over longer time horizons without stuffing the context full. The method focuses on history representation, not scaling up the model.
• Source: arXiv · 2026-11-04
• Editor's Take: The hardest part of a world model isn't generating the next frame; it's remembering the frame from thirty seconds ago. Pick the wrong compression strategy and the picture starts to drift.

3. Physical experience can feed a continuously updating spatial world model

• Summary: The paper discusses letting robots update their spatial world model through continuous interaction, rather than freezing it after training. The authors treat "physical experience" as a training signal, testing whether the model can still maintain spatial judgment after the environment changes.
• Source: arXiv · 2026-11-04
• Editor's Take: After deployment the environment changes: shelves move, floor tiles get replaced. A world model that can't update online starts expiring the day it ships.

4. Opinion: What physical AI lacks isn't a bigger brain, it's a nervous system

• Summary: This TechRadar commentary places the bottleneck of physical AI at the real-time feedback layer. The author argues compute and model scale are no longer the main constraints; the bandwidth between sensors, actuators, and control loops is, reasoning that systems that run well on real robots don't have large parameter counts.
• Source: TechRadar · 2026-11-04
• Editor's Take: This judgment holds up on the engineering side. Cram a 70B model into a robotic arm and response time collapses first.

---

4. AI Large Models

1. Anthropic open-sources FlashProve: using zero-knowledge proofs to say who said what

• Summary: Anthropic unveiled FlashProve, an open-source inference verification system that lets model providers cryptographically prove which model a given output came from. The GitHub version is 0.16.0, Apache 2.0, officially labeled a research artifact: unmaintained, no contributions accepted.
• Source: TechCrunch · 2026-11-04
• Editor's Take: Model routing, distillation, and wrappers all make money off information asymmetry. Once verifiable inference gets cheap, the middlemen's margins go first.

2. Quanta: Research tests whether models know what they're thinking

• Summary: The set of studies Quanta reports on tries to measure models' ability to report their own internal states, i.e., whether when it says "I was just considering X," that statement matches what actually happened internally.
• Source: Quanta Magazine · 2026-11-04
• Editor's Take: Interpretability wants to save costs by relying on models' self-reports, provided those reports are trustworthy. When a model says it doesn't know, you can't tell whether it didn't see or just won't say.

3. NVIDIA research: Agents actually rarely need long context

• Summary: This NVIDIA study, summarized by i10x, shows that after compressing the context of Claude Opus 4.5 and GPT 5.6 Sol, accuracy drops little or not at all, suggesting most agent tasks don't really depend on windows of hundreds of thousands of tokens.
• Source: i10x · 2026-11-04
• Editor's Take: Long context is a selling point, not a requirement. The most expensive stretch on the bill is often retrieval junk that can be compressed away.

4. StateBench: Testing whether models can track state changes in long documents

• Summary: AllenAI released StateBench on Hugging Face, isolating state tracking in long documents for testing. The task requires the model to remember that something was changed midway through the document, not just answer what was written at the start.
• Source: Hugging Face · 2026-11-04
• Editor's Take: Contracts, medical records, tickets—all long documents whose state changes. Retrieval augmentation often dredges up the old version here.

5. Hugging Face CEO: My p(doom) is close to zero

• Summary: At a WIRED event, Clément Delangue said he scores the probability of catastrophe close to zero. His reasoning: risk is more likely to come from a few big companies pushing ahead without disclosure, and open-source models serve as a defense here.
• Source: WIRED · 2026-11-05
• Editor's Take: This stance aligns with his business, which isn't entirely a bad thing. Open source does spread capability, and it also leaves regulators without a door to knock on.

6. ChatGPT can now generate animated GIFs directly

• Summary: The Verge reports that ChatGPT supports generating GIFs from prompts or uploaded photos, shipped with Images 2.6. Generated animations can be previewed and downloaded right in the conversation.
• Source: The Verge · 2026-11-05
• Editor's Take: Animated GIFs are the lowest-cost format for social sharing. What OpenAI is filling in is the sharing path, not image quality.

7. SambaNova and Supermicro co-build RDU + GPU inference systems

• Summary: HPCwire reports the two will put SambaNova's RDU and Supermicro's GPU servers into the same inference system, allocating chips by workload, with the pitch being inference cost rather than training.
• Source: HPCwire · 2026-11-05
• Editor's Take: The hard part of mixing is the scheduling layer. Customers buy cost per million tokens, not how many chip types are in the rack.

---

5. AI Software

1. Meta acquires AI coding company Conductor

• Summary: TechCrunch reports Meta bought Conductor. The company's product distributes multiple agents into parallel workspaces, letting users spin up a software project with one sentence, each running its own module.
• Source: TechCrunch · 2026-11-05
• Editor's Take: Parallel workspaces solve context contamination, not coding ability. Meta is buying this scheduling experience; the people may not stay.

2. Apple starts budgeting employees' AI usage

• Summary: Bloomberg reports Apple is cutting internal AI spending, setting quotas on AI calls for some employees and teams starting in October. MacRumors, relaying the story, said the goal is cost control, not limiting tool types.
• Source: Bloomberg · 2026-11-04
• Editor's Take: Even the company with the deepest cash flow is now counting tokens per head. The pricing inflection point in the enterprise market will come sooner than expected.

3. Goldman and Apollo in talks to resell SpaceX's $40 billion AI deal

• Summary: Bloomberg reports Goldman and Apollo are negotiating to resell this $40 billion AI deal tied to SpaceX to investors, structured closer to distribution of infrastructure assets than plain equity financing.
• Source: Bloomberg · 2026-11-05
• Editor's Take: Compute contracts are being turned into tradable assets. Risk moves from tech companies' balance sheets into funds' holdings.

4. Uber had 4 trips completed by AI agents last quarter

• Summary: On its earnings call, Uber disclosed that last quarter only 4 trips were ordered and completed via AI agents. The number came up in the section where the company discussed whether agents might bypass the platform.
• Source: The Verge · 2026-11-04
• Editor's Take: 4 trips. Every projection about agents disrupting entry points can first be checked against this number.

5. Google's K-12 policy draws criticism: restrict AI at school, lose Search

• Summary: TechCrunch reports Google announced that schools using its K-12 software would lose access to Search if they don't enable AI features like Gemini, drawing concentrated criticism from schools and parent groups after the news broke.
• Source: TechCrunch · 2026-11-05
• Editor's Take: Using Search as a bargaining chip shows Google weighs AI penetration in the education market more heavily than Search revenue.

6. AT&T's home security integrates Gemini for Home

• Summary: The Verge reports AT&T's Connected Life home security system is integrating Gemini for Home, letting users query camera and sensor status in natural language. AT&T calls it the first home security product to integrate the system.
• Source: The Verge · 2026-11-05
• Editor's Take: Security is one of the few scenarios where users will pay for "understanding." One fewer false alarm is worth more than ten extra features.

7. Cloudflare adds visibility and control for personal agents

• Summary: Cloudflare released new capabilities for personal agents, letting users see which sites an agent visited and what it did, and pre-set allowed and forbidden terms. The idea is to put permissions at the network layer rather than having each app implement its own.
• Source: Cloudflare Blog · 2026-11-04
• Editor's Take: Agent permissions are scattered across dozens of SaaS products; nobody can keep up. Putting them at the egress layer can at least block one.

8. Accenture contractor connected a personal Claude agent to Pentagon systems

• Summary: According to IBTimes, an Accenture contractor's personal Claude Code agent was alleged to have been connected to NIPRNet, the Pentagon network used by intelligence agencies. The report says the connection used a personal account, not a company-approved channel.
• Source: IBTimes · 2026-11-05
• Editor's Take: Shadow IT has a new form. It used to be employees installing software themselves; now it's employees wiring up agents themselves—and agents have credentials.

9. Kenyan startup builds AI on mobile payments and cheap phones

• Summary: The company covered by IEEE Spectrum doesn't build a general assistant; it connects AI to local mobile payment and small-merchant workflows, with low-cost Android phones as the terminal. The design premise is unstable networks and expensive data.
• Source: IEEE Spectrum · 2026-11-05
• Editor's Take: AI products in emerging markets face completely different constraints. Offline usability and per-message billing matter more than model scores.

---

Track counts: Humanoid robots 4 · Intelligent driving 3 · World models 4 · Large models 7 · AI software 9, 27 items total