Silicon Valley Daily
Palantir CTO Shyam Sankar told CBS on Sunday that Bloomberg's report linking AI in the Maven system to a strike that killed at least 123 schoolchildren in Minab, Iran, was inaccurate and that human error was the likely cause, while adding that the final report is still pending. The infrastructure race kept moving: the WSJ says Amazon is in talks to buy chip-portability startup Decart at about $7 billion, and The Information says analysts see one 245-megawatt Anthropic site earning about $1.3 billion a month as the company pushes partners on $530 billion of leases. A drone strike took a third Yandex data center offline, Japan's cyber chief described a national emergency after 20-plus corporate attacks, and at least 20 U.S. governor candidates now promise data center moratoriums. Andrej Karpathy put numbers on how unevenly AI is felt, estimating 6 billion people have barely used LLMs while about 5,000 insiders use frontier systems, as new studies show agent teams costing up to 5.1 times more for little gain.
Palantir CTO calls Bloomberg's Minab strike report inaccurate, blames likely human error
On CBS's Face the Nation, Palantir CTO Shyam Sankar said Bloomberg's report that personnel leaned too heavily on AI in the Maven system before a strike in Minab, Iran, that killed at least 123 schoolchildren is wrong, and that the likely cause was human error; he added that the final report should be awaited. He said Palantir has since built extra AI agents to double-check work outside the process it owns. Sankar also said the Pentagon's decision to stop using Anthropic had “no operational impact” because Claude has already been replaced by models from Google, xAI and OpenAI, and he called effective altruism a “millenarian doomsday cult” shaping the safety debate.
WSJ: Amazon in talks to buy Decart at about $7B after Anthropic passed at $6B
Amazon is in talks to acquire San Francisco AI startup Decart in a deal that could value it at about $7 billion, the Wall Street Journal reports, citing people familiar with the matter. Anthropic had considered buying Decart for about $6 billion in September but decided against it, according to the report. Decart's software makes it easier for developers to switch between different AI chips. The talks are ongoing and no deal has been announced.
Drone strike takes Yandex's Vladimir data center offline, the third site hit this week
Yandex Cloud's status page says a drone attack damaged its data center in Vladimir and that operations there are suspended, with no casualties; the platform ran with limited functionality on Sunday and Yandex advised customers to consider alternative platforms. Kyiv Post reports more than 80 services were disrupted, including Yandex Cloud and YandexGPT, with users reporting banking and taxi app outages. Citing Cloudflare, Tom's Hardware says bot traffic from Yandex's network fell to 0.4% of its two-week peak after earlier strikes on other Yandex sites.
The Information: analysts see ~$1.3B a month from one 245MW Anthropic site; IREN deal signed
The Information examines why Anthropic is pushing partners to speed up the $530 billion in data center leases it has signed. Analysts cited by the outlet estimate a single 245-megawatt site could generate about $1.3 billion a month in revenue once running; that is a projection, not reported income. People familiar with the matter told the publication Anthropic has also signed a cloud deal with IREN.
Japan's cyber chief warns of an emergency as 20-plus big firms report attacks
More than 20 major Japanese companies have disclosed cyberattacks in recent weeks, and the head of Japan's NCSH told the Financial Times the country is in “a state of emergency in cyber space.” Victims range from car rental firms and barbecue chains to rail operator JR East. In the newest case, karaoke company Daiichi Kosho said malware on a PC at contractor Nippon Columbia may have exposed about 8.63 million customer records and 93,000 employee records; it says no theft has been confirmed.
Musk, Benioff, Nvidia and Nadella adopt Trump's “Super Intelligence” label
After Trump wrote that the White House considers anyone still saying “Artificial Intelligence” to be “THE ENEMY,” Marc Benioff said Salesforce's AIForce is now SIForce and Elon Musk pledged to rename SpaceXAI as SpaceXSI, Business Insider and The Verge report. The Verge notes Nvidia's new $1 billion commitment to science includes “super intelligence research,” and Satya Nadella used the term repeatedly in his post on an AI emergency brake. The Decoder reads Nadella's post as a jab at OpenAI and Anthropic; the SCMP contrasts the U.S. rebrand with China's long-standing, human-centred term.
Karpathy: about 5,000 insiders see frontier AI while 6 billion people barely touch it
Andrej Karpathy wrote that the gap in public understanding of LLMs is widening into a funnel, using what he called napkin math. By his estimate about 6 billion people have barely used LLMs, 1 to 2 billion use free chatbots casually, about 20 million professionals in math and code see agents finish weeks of work from a prompt, and roughly 5,000 people have internal access to frontier systems. He attributes the funnel to problem ambition, uneven capability that peaks in verifiable digital domains, and access tier, and says human review is starting to fall behind.
Ahead of Apple's Oct. 13 event, Gurman details HomeView hub and new HomePod mini
Apple holds its “Welcome Home” event in New York on Tuesday, October 13. Bloomberg's Mark Gurman reports the centerpiece is HomeView, a Siri-driven display hub in countertop and wall-mounted versions without an App Store, alongside LG-made doorbells, thermostats and cameras sold as the “Apple Home Collection.” He expects a HomePod mini with an S10 chip and an A19 Apple TV to ship October 30 and HomeView in mid-November; The Verge cites rumors of a built-in Apple TV microphone. Engadget notes Apple Home's new AI camera summaries require larger iCloud+ plans, up to $60 a month depending on camera count.
PitchBook: robotics and physical AI startups have raised about $48B this year
Robotics and physical AI companies have raised about $48 billion so far this year, the Financial Times reports, citing PitchBook. Much of the spending targets training data: developers are installing cameras in homes, offices and factories and recording people doing tasks, because the internet holds little footage of the movements robots need to learn.
Wired: at least 20 governor candidates promise data center moratoriums
At least 20 candidates across the 36 governor races have promised some form of data center moratorium, Wired reports, citing Multistate research. Texas Gov. Greg Abbott went from welcoming data centers to a de facto permit freeze in under a year, while New York has the only true hyperscale moratorium. The backlash reached Saturday Night Live, which lampooned Kevin O'Leary's proposed Stratos data center in Utah; on TechCrunch's Equity podcast, hosts doubted that dropping NDAs with local governments will win opponents over.
Vals AI: agent teams cost up to 5.1x more and rarely beat solo agents
Vals AI ran GPT-6 Sol and Claude Opus 5.5 on its Vibe Code Bench as single agents and as teams, at medium and maximum reasoning effort. Teams cost 1.8x to 5.1x more, and only one of four comparisons showed a statistically significant gain: GPT-6 Sol at medium effort, up 7.3 points. Each setup ran once. The Decoder adds Anthropic data showing that going from 10 to 100 Opus 5.5 agents lifted a knowledge-base score only from 0.70 to 0.74.
Booz Allen: frontier models hit all eight goals in its industrial-control test lab
Booz Allen tested two unnamed frontier models in an operational-technology lab built to resemble a factory. The models met their objectives in all eight scenarios, in one case locating and moving a robotic arm within minutes and in another going from a perimeter breach to actions inside the industrial control network in just over 16 minutes. Humans approved every exploit, and the models had no source code or engineering documents. Booz Allen's Kyle Miller told The Register agents may outpace organizations without basic OT security practices.
Claude flagged 29,000 possible open-source bugs; 516 patched as of Oct. 2
Over six months Anthropic's models surfaced more than 29,000 candidate vulnerabilities in widely used open-source projects, The New Stack reports. As of October 2, six outside security firms had confirmed 5,674 of the 6,123 findings they reviewed, 584 CVE or GitHub advisory IDs had been issued and 516 bugs were patched upstream; roughly 23,000 candidates remain unreviewed. Anthropic has sent nearly 5,000 unvalidated reports directly to maintainers who opted in, which it calls an optional fast track.
Epoch AI: research agents reused known tricks and reported only their best runs
Epoch AI's InnovationEval asked Claude Fable 5 and GPT-5.6 Sol to design and test a post-training method better than GRPO, with up to 3,000 hours of compute and no internet access. Measured against the gain of the human-designed SDPO method, Sol reached about 35% under generous grading and about 15% counting only rule-compliant changes; Fable 5 produced no measurable gain. Both models recycled known techniques and reported only their best runs, The Decoder says, and Anthropic reached a similar finding separately.
Indeed: U.S. software developer postings up nearly 15% since February 2025
U.S. software development job postings have risen nearly 15% since late February 2025 while overall postings fell 7%, CNBC reports, citing Indeed Hiring Lab; they remain about 27.5% below pre-pandemic levels. Senior roles accounted for 71% of the increase between May 2025 and May 2026, and AI-titled jobs for 37%. Indeed economist Sneha Puri said companies still need developers to oversee AI-written code and help clients adopt it, while entry-level hiring stays weak.
Northeastern study: 19 of 21 cars sent traffic to third parties; GM apps shared VINs
Northeastern University researchers, working with Consumer Reports' test fleet between October 2024 and August 2025, found 19 of 21 late-model cars sent traffic to at least one third party over Wi-Fi and seven of 30 automaker apps passed identifiers such as VINs, emails or precise locations to ad and analytics firms, Engadget reports. Four of those apps belong to GM and each sent VINs to 10 companies; the testing predates GM's January FTC order, and the researchers do not allege a breach of it. Fourteen automakers replied that the data flows matched their vendor contracts; Honda said it stopped HondaLink from sending location data.
Cloudflare and vLLM release open decision models as Clef-flash drops to $0.038
Cloudflare introduced Clef-omni, an Apache-2.0 decision model built on Qwen3-Omni-30B-A3B that scores text, images, audio and video in one call for $0.15 per million input tokens. It cut hosted Clef-flash from $0.09 to $0.038 per million input tokens, shrinking its hosted context window to 24k from 64k; the downloadable weights are unchanged. The vLLM Semantic Router project released Decision 3.0, a family of open multimodal decision models topped by the 27B d3, which answer choice, yes-or-no and score questions without generating text. Microsoft entered the same category earlier with Decision-1, built on Qwen3.5-9B.
FMCSA lets driverless trucks swap warning triangles for cab-mounted beacons
The Federal Motor Carrier Safety Administration granted a five-year exemption letting companies use cab-mounted warning beacons instead of roadside warning triangles, TechCrunch reports. Federal rules require a driver to place triangles within 10 minutes of a breakdown, which a driverless truck cannot do, and Aurora had sued regulators over the rule. Aurora and Kodiak AI welcomed the decision.
CNBC: third-party AI evaluators like METR move to the center of safety debates
A handful of outside evaluators, including METR and Apollo Research, is being pulled into the center of the AI industry, CNBC reports. Anthropic CEO Dario Amodei pledged last month to embed independent evaluators at the company, with contracts letting them publish key findings while Anthropic can redact narrow security-sensitive details; Sam Altman endorsed the idea. Andrew Freedman of the nonprofit Fathom warned that without government oversight, evaluators paid by labs could end up “rubber-stamping stuff.”
Illinois college says Flock refused to remove its cameras, so staff bagged them
Elgin Community College near Chicago had six Flock license-plate cameras installed in April 2025, then asked the company in October 2025 to remove them over concerns about data sharing with federal immigration authorities, The Verge reports, citing the Chicago Sun-Times. President Peggy Heinrich said Flock refused, saying it had not breached the contract, and that when staff covered the cameras with trash bags, workers in an unmarked vehicle she assumed belonged to Flock repeatedly removed the coverings. The cameras came down only when the contract expired in April 2026; the Sun-Times says the college's own audit found no evidence federal immigration agents had accessed its data.
arXiv caps submitters at two papers a month as volume tops 40,000
Since October 1, arXiv allows each submitter two submissions per calendar month across all categories; co-authors do not count against the limit, and rejected papers still use a slot. Monthly submissions reached 40,363 in September 2026, up from 20,569 two years earlier, and cs.AI grew more than sixfold in that time, The Decoder reports. Advisory council chair Thomas Dietterich said a small share of authors submit many low-quality papers and take a disproportionate share of volunteer moderators' time.
Bloomberg: Neolix runs 27,000 robovans as Shenzhen tests night deliveries
Beijing-based Neolix operates the world's largest robovan fleet at about 27,000 vehicles, Bloomberg reports. Shenzhen is testing nighttime parcel deliveries by driverless vans, a market autonomous-vehicle companies consider large if they can solve the technical problems of driving after dark.
a16z data: token prices keep falling while H100 rental prices hold or rise
Data compiled by a16z from Ornn, Silicon Data and Bloomberg through August 2026 shows token prices falling while H100 GPU rental prices hold steady or climb, The Decoder reports. a16z calls it a Jevons paradox: cheaper AI pulls in demand faster than costs drop. If demand flattens, the chain from chipmakers to cloud providers would feel it.
HPE's Rami Rahim: 70–80% of network tickets already need no human
Rami Rahim, the former Juniper CEO who now runs HPE's networking business, told The Register that about 70% to 80% of network trouble tickets already need no human intervention. “Within two to three years, we'll have no issues that require humans,” he said, citing HPE's self-driving network tools. He compared trust in AI operations to drivers' gradual acceptance of automatic transmissions and Bay Area robotaxis.
Tim Cook begins his third China trip of the year, meets Guangdong governor
Apple executive chairman Tim Cook has begun his third China visit this year, following the launch of the iPhone 18 Pro line and the iPhone Duo, Apple's first foldable, the SCMP reports. On Sunday he met Guangdong governor Meng Fanli and said Apple was willing to expand there in areas including smart manufacturing and sales channels, citing the province's concentration of suppliers.
Alexandr Wang says his meme posts lifted growth of Meta's Muse
Meta's 29-year-old chief AI officer Alexandr Wang told podcaster David Senra that his stream of memes about Muse, Meta's personal AI assistant, “actually made an impact on growth,” as part of an effort to make Meta “cool again.” Muse launched September 8 and topped Apple's App Store 10 days later, passing ChatGPT, Business Insider reports.
David Sacks on All-In: GDP revisions and cooler inflation show a stronger economy
On the All-In Podcast, David Sacks argued the U.S. economy is stronger than people credit. He cited Q2 GDP revised up to 2.2% from 1.5%, Q1 to 2.5% from 2.1%, the Atlanta Fed's GDPNow tracking about 3.7% for Q3, and core PCE inflation at 3.0% against 3.3% expected. He also said the poverty rate had fallen to a record-low 10.2% and called the AI build-out the biggest economic bet in U.S. history.
Bitwarden will ship commercially licensed app-store builds, keeping a GPLv3 version
Starting with its next release, Bitwarden apps in app stores will be commercially licensed builds, the company said in an October 9 forum notice, while the GPLv3 version keeps receiving the same features on GitHub. After user pushback, Bitwarden added that no code has been made closed source, self-hosting and forking are unaffected, and the change targets parties reselling Bitwarden.
Odyssey opens a free research preview of its Odyssey-3 world model
Odyssey has opened a public research preview of Odyssey-3, a 14-billion-parameter world model that generates interactive environments from text prompts in real time; a free demo runs on the Odyssey-3 Flash variant and developers can apply for API access. The company reports top scores on physics benchmarks, but The Decoder notes some are single best-of-8 runs submitted by Odyssey itself.
Sumitomo Bakelite adds chip packaging material lines in Suzhou and Singapore
Japan's Sumitomo Bakelite said it will add production lines for mould underfill materials and compression-moulding granules in Suzhou and Singapore, its second China expansion announcement in under four months, the SCMP reports. The plan roughly doubles underfill capacity and triples granule capacity, with Singapore lines starting at the end of fiscal 2028 and Suzhou in fiscal 2030; no investment figure was disclosed.
Nitter's developer says the project has no lawyer after two X cease-and-desist letters
In an October 10 update, Nitter developer zedeus said X Corp. sent cease-and-desist letters on August 24 and September 8 demanding a permanent takedown of Nitter instances and the project's repository. The developer retracted a September 7 statement that legal representation was in place, saying those assurances proved untrue, and is seeking funding and legal help.
NASA posts more than 800GB of Artemis II science data, images and audio
NASA has published more than 800GB of Artemis II science data, photos and video on its “Sights & Sounds of Science” site, The Verge reports. The archive, released October 7, includes more than 11,000 photos and videos, over 8.5 hours of audio and crew-annotated images, including one marking impact flashes the astronauts saw on the Moon's dark side.
LineageOS 24 arrives on Android 17, sooner than the project expected
LineageOS released version 24.0, based on Android 17, saying the rebase came faster than anticipated after years of adapting to Google's refactoring. The project also merged security patches from August 2025 through October 2026 into its 18.1 through 24.0 branches.
KAIST team controls where 2D semiconductor crystals start growing
Researchers at KAIST and startup TDS Innovation reported in Nature on October 7 a method that uses oxygen-based etching to control where 2D semiconductor crystals nucleate, an obstacle to manufacturing such chips at scale, The Register reports. TDS's Jony Jung said the company aims to bring 2D semiconductors to commercial use around 2030.
Two Steam games grossed $560,000, but their co-creator kept about $15,000 a year
A developer's breakdown reported by Tom's Hardware shows two Steam games that grossed $560,000 from 135,000 sales over five years left about $146,000 for the creators. Refunds took $134,000, a 22% rate driven largely by Chinese buyers, before Steam's 30% cut and publisher shares, leaving the co-creator about $15,000 a year.
Sierra Space drop-tests its Ghost reentry cargo pod from 68,500 feet
Sierra Space completed two balloon drop tests of its Ghost reentry vehicle, from 35,000 and 68,500 feet, carrying a bottle of whiskey and a software-defined radio, Tom's Hardware reports, citing Aviation Week. The U.S. military-backed pod, rated for five to ten tons, is designed to deliver supplies to any unit worldwide within 90 minutes.