Community Discussion · General

WestTide · AI · 2026-10-21 · Issue 104

West TideOct 212026/10/20 274 views

Today's Briefing: The Wall Street Journal reports, and Bloomberg relays, that Anthropic could spend $750 billion over the next few years to expand compute. Meta's Muse hit 10.2 million downloads in the US within 40 days of launch, and in the same week paused chat sharing because shared links could be bulk-retrieved. Google open-sourced TorchTPU, moving the chip battle to the toolchain layer.

Editor's Note: Once the money reaches the hundred-billion scale, all that's left to discuss are the specific terms.

1. AI Large Models

1. Anthropic's compute bill could pile up to $750 billion

Summary: The Wall Street Journal reports, and Bloomberg relays, that Anthropic could spend $750 billion over the next few years to expand compute.
Source: Bloomberg · 2026-10-20
Editor's Take: Nobody looks at the total when the contract is signed. By the time they see the total, the term and default clauses are what matter.

2. ChatGPT can now lock chats with a PIN

Summary: OpenAI launched locked chats, requiring a four-digit PIN or Face ID to open. Locked conversations don't appear in history, pinned items, or search. The company notes this is not equivalent to end-to-end encryption and doesn't change model training settings or data retention periods; resetting the account PIN deletes locked conversations and uploaded files together, with no recovery. Free, Go, Plus, and Pro accounts are all in the rollout.
Source: Engadget · 2026-10-20
Editor's Take: It guards against the person next to you. The service provider can still see it.

3. A large model is going to run on a lunar rover

Summary: Astrolab's FLIP rover is set to depart as early as November, carrying an HPE high-performance computer in its cabin. It will run a custom large model on the lunar surface for the first time, analyzing camera and lidar data, assisting lunar navigation, and compressing the volume of data sent back to Earth. One-way Mars communication latency can reach 20 minutes, which is why autonomous decision-making is being pushed to the device.
Source: MIT Technology Review · 2026-10-20
Editor's Take: If you can't solve the latency, the model has to go with the device.

4. Gemini's "Personal Intelligence" comes to the car

Summary: Google is bringing Personal Intelligence into in-car Gemini, which can call on Gmail, Calendar, YouTube, Maps, and search history. Driving to the airport, a delayed flight in the calendar will proactively alert you; if the user keeps asking about the gate and the delay, they can get both answers at once. The feature launches first on Android Auto, with Google Built-in vehicles following. Users must turn it on themselves and can also choose to connect only some apps.
Source: TechCrunch · 2026-10-20
Editor's Take: The car is yet another entry point into the account ecosystem. Cross-app data stitching is the real substance of this update.

5. Early research by four Z.ai founders came from a Chinese institute the UK links to espionage

Summary: Bloomberg reports that four founders of Chinese large-model company Z.ai took part in research funded by a Chinese institute that the UK has linked to espionage.
Source: Bloomberg · 2026-10-19
Editor's Take: The landing point of this kind of reporting is compliance review checklists and financing inquiries, nothing to do with model scores.

2. AI Software

1. Meta pauses Muse's chat sharing feature

Summary: Security firm Claroc found that Muse's share links put the conversation title into the subdomain, and the addresses entered public DNS, so a single passive DNS query could enumerate them in bulk. Researchers enumerated 6,800 links, and the visible content included meeting times with a frontier AI lab, small-business loan applications, bank anti-fraud alerts with partial account info, and company vacation schedules. Meta responded that it has paused the feature and is evaluating link visibility rules.
Source: Engadget · 2026-10-20
Editor's Take: The domain rules weren't isolated, so sharing and public publishing are the same thing.

2. Meta's Muse hits 10 million US downloads in 40 days

Summary: Sensor Tower data shows Muse has accumulated 10.9 million downloads since launching in September, of which 10.2 million are in the US, making it the 12th app to reach that scale within 40 days. The last one was Threads in July 2023. In October, Meta directed 39% of its ad spend to Muse, up from 23% last month. US 7-day retention is 30% and 14-day retention is 23%, versus averages of 19% and 16% for ChatGPT, Claude, and Grok.
Source: TechCrunch · 2026-10-20
Editor's Take: Retention didn't collapse, so this round of ad spend was worth it.

3. Microsoft treats AI Agents as insider threats

Summary: Microsoft released Guardian Agents to keep watch on AI Agents running inside enterprises, under the Project Perception launched in June. Security business lead Hayete Gallot's line is that frontier models can't be trusted no matter where they're trained, not because of malice but because of unpredictability. Nadella previously wrote that enterprises need a system not controlled by model vendors to supervise Agents.
Source: TechCrunch · 2026-10-19
Editor's Take: Microsoft is an investor in both OpenAI and Anthropic. Coming from its mouth, this carries different weight than from other security vendors.

4. Google open-sources TorchTPU

Summary: Google announced Tuesday that its long-prepared PyTorch for TPU version, TorchTPU, is open as an open-source project. Together with Amazon's TorchNeuron, the two are competing for AI application developers at the toolchain level.
Source: TechCrunch · 2026-10-20
Editor's Take: The most expensive part of switching chips is migration. The toolchain is the first threshold deciding whether customers leave.

5. Otter.ai raises $320 million

Summary: The investor is General Catalyst's Customer Value Fund, which earns returns based on marketing results without diluting equity. Otter had previously raised about $50 million in total, crossed $100 million in ARR last year, and grew cumulative users from 25 million to 40 million. Monthly active users and churn rate were not disclosed. The funds go to the enterprise market, international expansion, and product iteration.
Source: TechCrunch · 2026-10-20
Editor's Take: The meeting-notes space is crowded. The numbers that would actually prove customers stay weren't given this time.

6. An AI Agent crashed an internet researchers' mailing list

Summary: On the second day of the AoIR annual conference, an AI identity calling itself Velvet Vale introduced itself on the 7,500-member Air-L mailing list and posed research questions about transparency and accountability for AI participants. Within nine minutes replies spiraled out of control. Dozens of researchers stated they hadn't consented to their emails being scraped, some demanded ethics review first, and some left the list. About an hour later the Agent was removed.
Source: Wired · 2026-10-20
Editor's Take: People who study the internet treated a public mailing list as their private living room.

7. OpenAI says its models can help banks migrate off COBOL

Summary: Engadget reports that OpenAI says its models perform better on migration tasks for legacy languages like COBOL, targeting banks whose core systems still run on old mainframes.
Source: Engadget · 2026-10-20
Editor's Take: The budget for this kind of overhaul has always existed. What's missing is a plan someone can sign off on.

3. Humanoid Robots

1. SoftBank's robot security business

Summary: Bloomberg reports that SoftBank is considering entering the robot security business, and a deal related to Allied Universal could be the entry point. Masayoshi Son has talked up robots repeatedly over the years; this time it lands on a security company's labor and sites.
Source: Bloomberg · 2026-10-20
Editor's Take: Security scenarios are fixed, billed per head, and the customers are institutions. The math is easier than home robots.

2. The "puffin" robot that flies and dives

Summary: A team from MIT and EPFL built a flapping-wing amphibious flyer called FAAV weighing under 300 grams, with hydrophobic nanocoating on the wings and an adjustable tail pitch angle, able to leap back into the air from water at a 70-degree pitch. Tests were done in a water tank and Lake Geneva. Next the team needs to tackle steering and performance in turbulent environments.
Source: MIT Technology Review · 2026-10-20
Editor's Take: For sampling waters traditional boats can't easily reach, the payers are marine science and nearshore monitoring.

3. A stroke rehab robot that learns from physical therapists using generative AI

Summary: A rehab system by Purdue's Johannes Lachner and Noah Geiger uses generative AI plus real-time force feedback to learn from real therapists' movements, giving appropriate assistance when patients do reaching, arm-raising, and similar training. The researchers stress the goal is to extend therapists' reach. The next phase is individualized models and long-term clinical controls.
Source: MIT Technology Review · 2026-10-20
Editor's Take: Millions are left with long-term impairments from stroke each year. The rehab gap is fundamentally a manpower gap, and the only thing that can be replicated is the model.

4. Nobody has settled the privacy bill for robot vacuums

Summary: Engadget lays out that connected vacuum models may upload cleaning records, floor plans, recognized objects, and room names. It cites one case where user Sammy Azdoufal wanted to remotely control his own DJI Romo but instead got control access to about 7,000 vacuums worldwide. DJI fixed the permission validation issue with two updates on February 10, 2026.
Source: Engadget · 2026-10-19
Editor's Take: Floor plans and voice going to the cloud is a feature, and also an exit for household information. When picking a model, the privacy terms deserve more attention than the suction specs.

4. Autonomous Driving

1. The physics lesson of Tesla's flying Roadster

Summary: WIRED breaks down last week's Roadster 2 prototype using a static friction model. The friction coefficient between rubber and asphalt is about 0.6, corresponding to a maximum climb angle of about 31 degrees. Even raising the coefficient to 0.9 only gets 42 degrees, yet the car in the presentation ran along a looped ramp to the point of being upside down. The article leaves whether humans should be allowed to drive on walls to lawyers and the FAA; physics has already computed the boundary.
Source: WIRED · 2026-10-20
Editor's Take: The presentation gives conclusions. Physics explains the conditions.

2. Kia trials remotely driven shared cars in Korea

Summary: Kia, together with Socar, KT, and autonomous driving startup SUM, is piloting Korea's first remotely driven shared cars in Pangyo Techno Valley. After users park a PV5 electric van at their destination, operators at the depot take over via 4G/5G and drive it back to the depot for charging and maintenance. The consortium built triple-redundant networks to handle disconnections, along with operator training and emergency procedures. Using existing communication infrastructure and mass-produced vehicles, the cost is lower than fully autonomous solutions.
Source: TechRadar · 2026-10-19
Editor's Take: Humans stay in the loop. What's saved is dispatch and parking. This path can be proven out first.

5. World Models / Physical AI

1. Eka valued at $1.45 billion, betting on force sensing

Summary: Eka Robotics closed a $150 million Series A at a post-money valuation of $1.45 billion, led by a16z and F-Prime. The founders are MIT professor Pulkit Agrawal and former DeepMind researcher Tuomas Haarnoja. The company takes a vision-force-action route with force-sensing grippers, claiming it can grasp 1,500 objects under 1 kg per hour with reliability above 95%, versus roughly 600 for humans.
Source: TechCrunch · 2026-10-20
Editor's Take: Grasping has always been bottlenecked by data. Force signals let the model stop relying only on the picture.

2. NVIDIA moves surgery into generative simulation

Summary: NVIDIA released Cosmos3-H-Dreams, based on the 16-billion-parameter Cosmos3-H-Embodied-Nano, trained on video and kinematics from gallbladder removal surgery. It runs interactively on a single RTX PRO 6000 and supports 28 action dimensions covering dual-arm pose, grippers, and energy control. The company states this is a research simulator, not a validated clinical system.
Source: Hugging Face · 2026-10-21
Editor's Take: Skipping geometric modeling and learning visual dynamics directly from video, this route is now probing high-difficulty scenarios like surgery.

3. Interpretability placed at the heart of physical AI

Summary: The Information published an interview with World Mechanics CEO Sonia Joseph, who argues interpretability is the key problem for physical AI.
Source: The Information · 2026-10-20
Editor's Take: When a robot errs, you need to pinpoint the exact step. A black box is harder to account for in the physical world than in a chat window.

Track tally: Large models 5 · Software 7 · Humanoid robots 4 · Autonomous driving 2 · World models/Physical AI 3, 21 total.

General

1. AI Large Models

2. AI Software

3. Humanoid Robots

4. Autonomous Driving

5. World Models / Physical AI

2 replies