Drones hit three Yandex data centers in four days, and AI compute is now a target

😁 Hello, super humans! Tuesday’s big story is not about a model or a benchmark, it is about the buildings the models live in. Three data centers belonging to one of Russia’s biggest AI and cloud companies went down in four days, and the lesson for anyone running infrastructure is uncomfortably practical. We also have Python 3.15, a new boss at Boston Dynamics, and a very big rack from AMD.

📰 Quick Signals

• 🧠 AI: A White House AI task force told AI companies that incident notification and remediation is “not optional,” after Anthropic disclosed that test models misused government web forms; the rules apply to all AI companies, with no deadlines or penalties stated.
• 🤖 Robotics: Boston Dynamics named Rohit Prasad, formerly Amazon’s head scientist for Alexa and AGI, as CEO effective October 7, with a focus on commercializing physical AI.
• 💻 Programming: Python 3.15 shipped on October 9 with built-in frozendict and sentinel types, explicit lazy imports, UTF-8 as the default encoding, and an upgraded experimental JIT.
• ⚡ Electronics: AMD unveiled Helion at the OCP Global Summit, a rack for agentic AI built on 6th Gen EPYC “Venice” CPUs with up to 32,768 cores, with partner configurations expected in November.
• 📡 Telecom: Nokia has dropped new custom silicon with Marvell and will ship Nvidia-GPU-based RAN products, with operator trials this quarter and commercial availability expected early next year.

🔍 The Big Story: Drones hit three Yandex data centers in four days, and AI compute is now a target

If your AI stack depends on one company, one region, or one campus, this week showed what a physical outage looks like. It is also a reminder that “the cloud” is a set of buildings with addresses.

What happened: According to Al Jazeera, Yandex said its data center in Vladimir, east of Moscow, was shut down after a drone attack, with no injuries reported. The company said operations there were completely halted, told customers to activate their disaster-recovery plans, and said the platform was in emergency mode with a remaining configuration “considered unstable.” It was the third attack on Yandex data centers in a week: Sasovo on October 8, the Kaluga region on October 9, and Vladimir on October 11.

The details: The Sasovo hub is the one that matters most for AI. Al Jazeera reports it houses two of the three supercomputers Yandex uses to develop its AI model, plus tens of thousands of servers, and Cybernews notes it is unclear whether those machines were damaged. Yandex Cloud is still running with limited functionality, so customers who built on it are now living through a real failover test. Al Jazeera does not explicitly attribute the Vladimir strike; it notes that Ukrainian President Zelenskyy said last week that Ukraine is responding to strikes on its own data centers, and that Russia has hit Ukrainian data centers and telecom facilities, including sites in central Kyiv, over the past month. Cybernews adds that the Kremlin has ordered data center operators to prepare for further drone attacks.

flowchart LR
    A["Oct 8: Sasovo hub shut down by a drone strike"] --> B["Oct 9: Kaluga region facility partly knocked out"]
    B --> C["Oct 11: Vladimir data center fully halted"]
    A -.houses.-> S["2 of 3 supercomputers used to train Yandex's AI model"]
    C --> D["Yandex Cloud in emergency mode; customers told to use DR plans"]

🗞️ More News

🧠 AI

• Anthropic opened a free, opt-in vulnerability-scanning service for open-source maintainers; it says it found more than 29,000 candidate vulnerabilities over six months, and in a check of 97 critical and high-severity findings across 48 projects, 85 met the bar for its disclosure process and one was a false positive.
• Microsoft detailed Maia 300, its third-generation AI accelerator: TSMC N2P compute, eight HBM4 stacks for 288 GB and 18 TB/s per chip, and a 167 kW liquid-cooled rack of 96 accelerators; Microsoft claims up to 7x the rack-scale token throughput of Maia 200, which is unverified.
• A survey of 300 senior engineering leaders by debugging vendor Undo found 81% had a production incident linked to AI coding tools in the past six months, 93% said an AI hallucination led to a misdiagnosed root cause, and 79% said release cycles are no faster than before AI.
• The Verge’s hands-on test of running AI locally, with the open-source Hermes Agent app and a 125B-parameter model on a 256GB Mac Studio, found the setup overwhelming at times, with model-choice overload and morning-briefing automations that break.
• Business Insider reports hyperscalers are funding free worker meals at data-center build sites, as tech giants are projected to spend over $5 trillion on AI and data centers through 2030.
• Apple’s smart home hub will reportedly be called HomeView and ship in mid-November, with an A18 chip and a roughly 6-inch square screen, according to Bloomberg’s Mark Gurman as reported by MacRumors.

🤖 Robotics

• Agility Robotics and FORT Robotics signed a memorandum of understanding to extend FORT’s safety architecture to the Digit 5 humanoid, covering a safety pendant, on-robot communications, and an off-robot “Offboard Safety Bridge.”
• The Robot Report argues “boring” robots will win in the real world, and that robot form should follow the workflow instead of chasing social media attention.
• AVIDrone won the DARPA Lift Challenge with an electric helicopter that lifted 3.84 times its dry weight, 112 lb of payload, ahead of a battery-powered hexrotor at 3.64:1.

💻 Programming

• Linux 7.3-rc7 is out; Linus Torvalds said he would have preferred a smaller rc but the final release is still planned for next weekend unless something new comes up.
• The ext4 data=journal mode was deprecated in mainline this week and is scheduled for removal at the beginning of 2028.
• Python 3.15 adds the Tachyon sampling profiler, which attaches to running processes and can output flame graphs, alongside a new profiling package.
• Stable kernels 7.2.10 and 6.18.56 shipped on October 11, each with several hundred fixes.

⚡ Electronics

• Applied Materials opened its $5 billion EPIC Center in Silicon Valley, with TSMC, Samsung, SK Hynix, Micron, Intel and Broadcom as partners, aiming to shorten the path from chip research to production yields.
• A developer showed that the MediaTek MT7996 radio in the Gemtek W1700K WiFi 7 router can be used as a software-defined radio receiver across the 2.4, 5, and 6 GHz bands.
• EE Times looks at Europe’s advanced packaging gap, where Fraunhofer and Swissbit are targeting low-volume production between research labs and mass manufacturing.

📡 Telecom

• Ookla’s RootMetrics found 4G VoLTE still sets up calls faster than 5G VoNR in many places; Seoul led with a 1.06 second median VoLTE setup, while six of eight carriers tested were faster on VoNR than VoLTE.
• Amazon Leo has deployed 396 broadband satellites over 14 launches and says it can now build 27 satellites a week, but it missed its mid-2026 service target and still has no firm launch date.
• A judge has halted the Cable One and Mega Broadband deal just before it was set to close.
• Telia is making another round of job cuts.

👨‍💻 Code Corner

Python 3.15’s new frozendict is an immutable, hashable mapping, which makes it a good fit for configuration like a failover order that nothing should mutate at runtime. This sketch picks the first healthy site from a fixed priority list.

# failover.py: needs Python 3.15 (frozendict is a new built-in)
SITES = frozendict(primary="site-a", secondary="site-b", cold="site-c")

def pick_site(healthy: set[str]) -> str:
    for name in SITES.values():  # insertion order is preserved
        if name in healthy:
            return name
    raise RuntimeError("no healthy site: time to call someone")

print(pick_site({"site-b", "site-c"}))   # site-b
cache = {SITES: "config-v1"}             # hashable, so it can be a dict key
print(cache[SITES])                      # config-v1

Because the mapping cannot be changed after creation, a bug elsewhere cannot silently reorder your failover plan.

🧰 Toolbox

• rea: An open-source framework that lets AI agents analyze software from runtime behavior down to native compiled binaries, aimed at reverse engineering and auditing closed-source programs.
• context-mode: Sandboxes tool outputs to cut AI coding agent context use (claimed at 98%), adds persistent session memory, and supports 17 platforms.
• diagram-design: A diagram system for AI coding assistants with 44 diagram types built as standalone HTML and SVG, meant to replace generic Mermaid output.
• w1700k-sdr: Code that streams raw radio data from a Gemtek W1700K WiFi 7 router to a Linux machine, turning a router into a receive-only SDR.
• OpenScreen 2: A free, MIT-licensed screen recorder for Windows, macOS and Linux with local Whisper transcription and a local MCP server so coding agents can edit your video.

🎬 Demo Watch (rotating)

Sakana AI’s MASS: A single open-weights model (Qwen3.6-27B) plays three roles at once, task solver, workflow optimizer and evaluator of its own output, with no external supervisor. The authors report better score per output token than the untouched base model on research-style benchmarks, about 1.60x on MLR-Bench and 1.35x on DSBench, and a head-to-head win rate that rose from 53.9% to 69.9% after a second round. The catch: according to a summary of the work, it did slightly worse than the base model on coding-heavy benchmarks such as SWE-bench Verified, and the authors themselves flag the shared blind spot of a loop where one model judges itself as “an Achilles’ heel of recursive self-improvement.” Real signal: self-improvement without a teacher is working on some tasks. Hype to ignore: any claim that this generalizes to everything.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: Episode 3 of the camera-to-AI series, picking up after clean pixels left the camera and crossed the network.
• Building Your First Neuron From Scratch: Part 2 of the neural network basics, on how a network is a stack of simple, learnable transformations trained by nudging weights.
• The Network Behind the Cameras: The unglamorous plumbing that moves data from the camera to the model, a nice companion to today’s story about how much depends on infrastructure staying up.

😀 The Bot Says…

Yandex told its customers to activate their disaster-recovery plans, which is the most polite way I have ever seen to say “we hope you tested that.”

---

That’s all for today! Hit reply and tell us: when did you last actually fail over your own backup, and did it work?