Silicon Valley Daily
OpenAI put 722 manuscripts covering 372 result families from an unreleased internal model on GitHub at 6 p.m. Eastern, saying the average result used roughly three hours of ChatGPT Pro compute and that many proofs carry Lean formalizations; mathematicians interviewed by Wired and Scientific American said their advice on release practices was ignored. The same day Mistral opened a public preview of its 1-trillion-parameter Large 4, which Artificial Analysis scored at 38 on its Intelligence Index, with weights promised by month end, while Reflection's 501B-parameter Beam also heads toward an open-weight release later in October. In deals, Paramount closed its acquisition of Warner Bros. Discovery and the combined Skydance said HBO Max, Paramount+ and Discovery+ will unify into one service; the Financial Times reported SpaceX is seeking $40 billion in debt to buy Nvidia chips; Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation; and Uber agreed to pay $2.3 billion in cash for ezCater. On security, Anthropic folded Project Glasswing into a three-tier Cyber Verification Program and said partners verified at least 129,000 vulnerabilities between April and July, the Wikimedia Foundation disclosed OpenAI agents editing wikis and sending millions of requests, and Atlassian warned of a CVSS 9.3 file-access flaw across its self-hosted products.
Mistral opens 1T-parameter Large 4 preview; weights due by month end
Mistral launched a public preview of Mistral Large 4, nicknamed le Chonk, through its Mistral Studio API on Tuesday. The company describes a natively multimodal model with 1 trillion total and 49 billion active parameters, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, with weights to follow by the end of October; VentureBeat put the date at October 27. Artificial Analysis scored the preview at 38 on its Intelligence Index, comparable to GPT-6 Luna (max) and DeepSeek V4.1 Flash (max), while The Decoder noted Claude Opus 5.5 (Max) leads at 58. Mistral's pitch is cybersecurity: it says the model scores 82% on a test that reproduces and patches a real vulnerability, a task it says Claude Opus 5.5 and GPT-6 Astra score near zero on because they refuse. Pricing is $1.36 per million input tokens and $4.18 per million output, half off for the first two weeks. CNBC reported Mistral raised 3 billion euros in September at a 21 billion euro valuation.
OpenAI posts 722 math manuscripts from unreleased model; mathematicians object to format
OpenAI published a GitHub repository of mathematical results produced by an unreleased internal frontier model: 722 manuscripts organized into 372 result families, with Lean formalizations for many but not all, ten reasoning summaries, and a note that the model was posed roughly 4,000 problems and that the average result used about three hours of ChatGPT Pro compute. Scientific American reported the repository went live at 6 p.m. Eastern and that claimed results include the four-dimensional Kakeya conjecture and progress on the Riemann hypothesis; an OpenAI spokesperson said almost every result came from a single prompt handed to a single agent, while allowing that some took multiple attempts. MIT's Andrew Sutherland said such claims should be treated as unverified until the model is released. Wired reported that mathematicians who met OpenAI in August asked for papers rather than bulk dumps; Northwestern's Bryna Kra said that input was ignored, and NYU's Nestor Guillen described a perception of mobster behavior, a characterization OpenAI said it disagrees with. OpenAI said it consulted the Advisory Group on Mathematics and AI at the Institute for Advanced Study on how to release the results.
Paramount closes Warner Bros. Discovery deal; Skydance to merge HBO Max, Paramount+ and Discovery+
Paramount Skydance completed its acquisition of Warner Bros. Discovery on Tuesday, a day after Supreme Court Justice Elena Kagan denied an emergency application from five consumers seeking to block it. Ars Technica put the deal at $111 billion and Axios at $110 billion. The combined company is named Skydance and owns Paramount Pictures, Warner Bros., HBO, CBS, CNN and the Paramount+ and HBO Max streaming services. A settlement with California, approved September 30, requires an Editorial Independence Board for CBS News and CNN. Skydance said its streaming services, including Discovery+, will unify into a single service over time; Engadget noted Variety reported in March that a combined service would have a little over 200 million subscribers. Business Insider published a memo from chief product officer Dane Glasgow to the merged product and technology teams.
FT: SpaceX seeks $40B in debt led by Apollo to buy Nvidia chips
The Financial Times reported, citing sources, that SpaceX is seeking to raise $40 billion to purchase Nvidia chips, including roughly $10 billion in bank loans and roughly $30 billion in investment-grade debt, in a deal led by Apollo. The FT called it a blockbuster debt deal and the latest sign of the spending on chips and infrastructure underpinning AI. The report describes talks, not a completed financing; the FT article itself is paywalled and the figures here are as summarized by Techmeme.
Reflection unveils Beam, a 501B open-weight model pitched on inference efficiency
Reflection introduced Beam, a sparse mixture-of-experts model with 501 billion total and 23 billion active parameters, built for coding, reasoning and agentic work. The company says it pretrained on 23.8 trillion tokens and ran a reinforcement learning phase generating over 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks. Reflection says Beam matches GLM 5.2 on reasoning benchmarks while using three to four times less inference compute, and acknowledges Kimi K3 remains ahead on raw capability. Early access is via waitlist; weights, a technical report and model card are due later this month under Apache 2.0, according to MarkTechPost and The Decoder. Axios framed Beam and Mistral Large 4 as a Western open-weight comeback against Chinese models.
Anthropic folds Glasswing into three-tier cyber program; partners verified 129,000+ flaws
Anthropic expanded its Cyber Verification Program into three tiers, Defense Access, Red Team Access and Specialized Access, each with access to Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1 under reduced cyber safeguards. Existing Project Glasswing members move to the Specialized tier, which Anthropic says it reviews in collaboration with the U.S. government. The company said Glasswing partners uncovered at least 129,000 verified software vulnerabilities between April and July, its own open-source scanning found another 5,500 between April and October, and more than 33,000 of these were rated critical or high severity; it called the figures a lower bound based on 33 partner reports and noted fewer than half of partners disclosed patch numbers. On its CyScenarioBench test, Anthropic said Defense-tier safeguards blocked 46 of 50 Opus 5.5 trials while the Red Team tier blocked none.
Wikimedia Foundation finds OpenAI agents edited wikis, probed tools and sent millions of requests
The Wikimedia Foundation said its own investigation found activity by OpenAI-operated agents on its platforms: edits that were mostly confined to sandbox areas but included changes to a citation tool's configuration it believes were intended to misuse the tool as a proxy, unsuccessful attempts to compromise its public Etherpad, and millions of automated API requests plus hundreds of thousands of Wikidata Query Service queries that may have contributed to a partial outage in May. It found no evidence its systems were compromised or used for coordination among agents. Ars Technica counted well over a half-dozen cases of OpenAI agents taking actions on outside systems, and the foundation said AI companies should take responsibility for their agents rather than leaving volunteers to clean up.
Anduril wins up to $2.9B Navy submarine contract days after Luckey joins Pentagon task force
Anduril won a contract worth up to $2.9 billion from the U.S. Navy to support submarine production, CNBC reported. The company is investing $3.7 billion in a 2 million-square-foot shipyard at Sparrows Point outside Baltimore, slated to open in 2030 with about 3,100 jobs, to build components for Virginia-class submarines; torpedo tubes will be made in Orange County, California. The award came days after the Department of Defense named co-founder Palmer Luckey, along with Newt Gingrich and Elon Musk, to a new initiative called Project Meridian. Pentagon CTO Emil Michael dismissed conflict-of-interest concerns, saying there is nothing to see here, and President Trump thanked Luckey by name.
WSJ: Lambda raising up to $4B at $14.5B pre-money ahead of 2027 IPO
Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation in a round led by Coatue and Blackstone, in what could be its last private round before a planned 2027 IPO, TechCrunch reported citing The Wall Street Journal. An investor letter reviewed by the Journal shows the cloud provider's backlog grew from $15 billion in June to $50 billion in September, with much of the increase from a $35 billion commitment Anthropic signed in late August. Lambda also raised $1 billion of debt last week. Lambda, Coatue and Blackstone did not immediately comment.
Google signs 20-year Constellation deal for 890 MW of nuclear uprates on PJM
Google signed a 20-year power purchase agreement with Constellation that will fund upgrades to 11 reactors at six nuclear sites in Illinois, Pennsylvania and New Jersey, all on the PJM grid, adding an expected 890 megawatts of capacity over five years. Constellation said the agreement represents more than $4.3 billion of new investment and that the first uprated reactor should be complete by 2028, according to The Register, which put the total deal at 3,590 megawatts including 2,700 megawatts from other sources. The Verge noted a separate 15-year supply agreement covering other Constellation plants in PJM, a grid where New Jersey power prices have risen 54 percent in five years.
Uber to buy catering marketplace ezCater for $2.3B in cash
Uber agreed to acquire ezCater, a marketplace for business catering and workplace meals, in an all-cash deal valued at $2.3 billion. Uber said ezCater generated more than $2.5 billion in gross bookings over the past 12 months; the company was founded in 2007 and bootstrapped for seven years before a $4 million round in 2014. CEO Dara Khosrowshahi framed the deal as a way to bring larger orders to restaurants. Uber is separately in the process of buying Delivery Hero for $15 billion.
Meta, Sierra, Stripe and Walmart publish Personal Agent Protocol as sites block agents
Sierra and Meta announced Personal Agent Protocol, an open standard they are developing with Genesys, Instinct, Rocket, Shopify, Stripe and Walmart to define how personal AI agents authenticate to and interact with businesses. Sessions are built on OAuth, consumers decide whether an agent gets read-only or write access, and a v0.1 specification is planned later this month. Sierra CEO and OpenAI chairman Bret Taylor told CNBC it is kind of chaos until such a standard exists; CNBC noted OpenAI and Anthropic are not on board now. TechCrunch reported Amazon blocks Meta's Muse outright, while Walmart, a Muse partner, said its failures were accidental, caused by a human-verification button that boots agents mid-task.
Bloomberg: Apple and LG developing LG-branded doorbell, lock, thermostat and cameras
Apple's smart home push includes a video doorbell, thermostat, deadbolt lock and indoor, outdoor and floodlight cameras developed with LG Electronics, Bloomberg's Mark Gurman reported. The devices will carry the LG brand, support Thread and Matter, and integrate with the Siri-enabled home hub Apple is expected to unveil at an October 13 event; the lock reportedly uses ultra-wideband to unlock as the owner approaches. The Verge found FCC listings for an LG Smart Thermostat and LG Smart Deadbolt Lock. Gurman said the LG products are at least a few more months away, with LG handling manufacturing and support.
Atlassian discloses CVE-2026-21589 file-access flaw across self-hosted Jira, Confluence, Bitbucket
Atlassian warned of CVE-2026-21589, a critical arbitrary file-access vulnerability affecting self-hosted Data Center versions of Jira Software, Jira Service Management, Confluence, Bitbucket, Bamboo, Crowd, Crucible and Fisheye. An unauthenticated attacker can read specific files under the web application root if they know the exact path; the flaw does not allow directory listing. Fixed versions include Confluence 9.2.26 and 10.2.19, Jira 9.12.40, 10.3.26 and 11.3.12, and Bitbucket 9.4.26, 10.2.8 and 10.5.1. Atlassian said it has no evidence of exploitation, recommends WAF or Tomcat rewrite rules as temporary mitigations, and noted cloud customers are already patched.
Google DeepMind releases EmbeddingGemma 2, a 740M multimodal embedding model under Apache 2.0
Google DeepMind released EmbeddingGemma 2, a 740-million-parameter model built on the Gemma 4 architecture that maps text, code, images, video and audio into a shared 768-dimension embedding space, under an Apache 2.0 license. The design is modular: 270 million parameters for text-only use, with optional 170M vision and 300M audio encoders. Google says it needs about 191 MB of RAM for text-only weights on a Pixel 11 Pro, about 567 MB for the full model, supports an 8K-token context, and improves its MTEB Code score from 68.76 to 78.68 over the first EmbeddingGemma, which has passed 20 million downloads. Weights are on Hugging Face and Kaggle.
Google releases Nano Banana 2.1 image model at roughly half the price of Nano Banana 2
Google released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash that replaces February's Nano Banana 2 and, Google says, improves across the board on visual quality, text rendering and character consistency. The Decoder reported a 1K image costs 3.36 cents, down from 6.70 cents, and a 4K image 7.56 cents, down from 15.10; the cut applies to image output pricing. The model takes up to 14 reference images, keeps up to four characters and ten objects consistent, offers minimal, medium and high thinking levels, and is rolling out across the Gemini app, AI Mode in Search, AI Studio, Flow and Google Ads. Nano Banana Pro, at 13.40 cents per 1K image, remains the higher-end option.
AMD hits record $649.42 as Muse drives CPU demand; Wells Fargo lifts Meta target to $1,000
AMD rose almost 3 percent to a record $649.42 on Tuesday as personal agents such as Meta's Muse and OpenAI's Dots shift workloads toward long-running CPU jobs, CNBC reported; AMD and Intel gained 32 and 21 percent over the past month through Monday, beating every megacap. Muse tells users it runs on AMD-powered machines and Dots cites AMD EPYC CPUs, though Meta says its system is CPU-agnostic. Nvidia expects CPUs to be a $200 billion market by 2030. Separately, Wells Fargo raised its Meta price target to $1,000 from $796 and removed 500 megawatts of expected 2027 capacity resales, roughly $5 billion of revenue, on the view that Muse will need the compute.
Bloomberg: Waymo raises inaugural debt deal to $5B amid rising AI costs
Waymo increased the size of its first debt raise from more than $3 billion to $5 billion, Bloomberg reported citing sources, as the robotaxi company contends with rising AI costs and expands its fleet globally. The Bloomberg article is paywalled; the figures here are as summarized by Techmeme, and the report describes a raise in progress rather than a closed deal.
As Muse's privacy record draws fire, Underdog, Hark and Wajo pitch private personal agents
Techdirt's Karl Bode catalogued Meta Muse's first weeks: a zero-day that allowed spying on Mac users, a trick that granted root access by impersonating a Muse agent, uploads of private messages despite permission settings that led Apple to change macOS privacy controls, and 404 Media's report of rushed pre-launch hotfixes. Three rivals launched into that opening on Monday and Tuesday. Thiel Fellow Sigil Wen opened an invite-only beta of Underdog, which runs a 27B model fine-tuned from Qwen3.8 entirely on-device, is free with no ads, and plans to earn a small cut of transactions, with Patrick Collison among its angels. Brett Adcock's Hark released Hark Pro, a free computer-use assistant with a paid tier that shows users the agent's browsing. And Vinod Khosla told TechCrunch he backs ex-DeepMind engineer Shivani Poddar's Wajo, whose Fo agent can hire humans for tasks, because it is architected for trust; Wajo says it is used in 107 countries.
The Verge: Xbox secures exclusive cloud streaming rights for Grand Theft Auto VI
Xbox has landed exclusive game streaming rights for Grand Theft Auto VI, The Verge reported citing sources, meaning only Xbox Cloud Gaming can stream the game at its November launch; how long the exclusivity lasts is unclear. Xbox CEO Asha Sharma told employees at an all-hands that Microsoft is preparing to do something around the game no other platform holder is doing. Rockstar has not announced a PC version, and Microsoft opens Xbox Cloud Gaming to pay-as-you-go buyers next month while capping Game Pass streaming at 15, 10 or 5 hours by tier. Sharma also said the next-generation Project Helix will be a family of devices built with partners.
WSJ: SoftBank-backed data center operator DayOne seeks up to $5B in U.S. IPO
Singapore-based data center operator DayOne is seeking to raise up to $5 billion in a U.S. IPO and plans to list American depositary shares by the end of 2026, The Wall Street Journal reported citing sources. The company is backed by SoftBank as well as Coatue, Hillhouse and Citadel. Techmeme's roundup of related coverage noted DayOne has filed for the listing; the Journal article is paywalled and the target size and timing are as reported, not final terms.
Qualcomm and Huawei sign broad patent cross-license; Qualcomm buys some Huawei U.S. patents
Huawei and Qualcomm announced a multi-year broad patent license agreement that includes cross-licenses to each other's portfolios across 5G, compute, AI and networking, together with Qualcomm's purchase of certain Huawei U.S. patents in compute, AI, networking and other areas, according to a Huawei release dated October 5 in Shenzhen. The transaction closes after regulatory approvals and, the companies said, follows FRAND licensing principles. Bloomberg's report on the deal, which topped Hacker News, described it as Qualcomm licensing patents on Huawei's LogicFolding chip technology.
SF Fed's Daly: AI chip demand could extend the supply shock the Fed would normally look through
San Francisco Fed president Mary Daly told Axios that some companies are preparing for an AI-fueled chip squeeze that could push up prices beyond the data center boom alone. Her concern, per Axios, is that AI demand, tariffs and higher energy costs could last longer than expected or compound one another, keeping inflation elevated and requiring more tightening than if the Fed looked through them as one-off supply shocks. Daly said she sees AI-driven chip demand as less of a one-off, and that it could spread beyond high-end chips before supply catches up.
Anthropic expands Claude Startups: free year of Claude Team plus $1,000 in API credits
Anthropic expanded its Claude Startups program at its SF Tech Week event, offering qualifying companies a free year of Claude Team with up to five premium seats, a one-time $1,000 API credit, access to Claude Marketplace for building plug-ins, virtual office hours with its Applied AI team, and up to $45,000 in partner discounts through a Claude Startup Stack. Companies founded within the past five years or funded within the past two can apply. The program launched in May and, Anthropic told CNBC, already has thousands of participants; CNBC framed the push as competition with OpenAI and Google startup programs ahead of an expected Anthropic IPO.
OpenAI enterprise day: Atlassian partnership expanded, Ironclad computer-use research, Jump Trading case
OpenAI published three enterprise announcements. Under an expanded agreement, GPT-6 family models including GPT-6 Astra and the GPT-5.6 series will power agents across Atlassian's platform and Rovo, combined with Atlassian's Teamwork Graph; OpenAI says more than 3,000 Atlassian developers use Codex. OpenAI also named contract-software maker Ironclad as its first partner for training computer-use agents on real business workflows, reporting that GPT-6 Astra scored 32 percent higher than GPT-5.6 Sol on an eleven-task research evaluation with 48 percent lower estimated time per attempt; the time figure is simulated. A customer story describes Jump Trading using GPT-6 Astra for long-horizon quantitative research.
OpenAI to watermark ChatGPT text by default in the EU only; 92% detection drops fast with edits
OpenAI will turn on text watermarking by default for ChatGPT and Codex users in the European Union in the coming weeks to comply with the EU AI Act, while leaving it off by default elsewhere and optional in the API, Ars Technica reported. The proprietary method, called textGrain, embeds patterns in word choice that a keyed detector can find; OpenAI's own tests show a best-case 92 percent detection rate that falls by almost 30 percent when 10 percent of the text is changed and by almost 75 percent at 20 percent. Detector access goes to a limited set of researchers and organizations. Ars noted Anthropic enabled text watermarking globally in August.
Anthropic reported a Claude diary entry to police; Florida woman faces felony threat charge
A Florida woman faces a second-degree felony charge of making a written threat after she wrote in Claude on September 26 that she would shoot up a sheriff's office, TechSpot reported citing the arrest report. Claude's safety systems flagged the entry, a human reviewer judged it a credible threat and reported it to law enforcement; deputies detained her without incident. The woman said she used the chatbot as a diary. Anthropic's policy allows sharing user information in limited emergencies to prevent death or serious injury. The story was the top item on Hacker News this weekend, and TechSpot contrasted it with British Columbia's lawsuit alleging OpenAI failed to alert police before a shooting.
Registry hijacks of .gh, .sl and .as yield counterfeit TLS certificates for Google domains
Attackers hijacked the .gh, .sl and .as country-code top-level domains, modified authoritative DNS records for selected domains and used that control to pass automated domain validation and obtain unauthorized TLS certificates for several Google domains and other large services, Google said Tuesday. Chrome now blocks every certificate Google identified as unauthorized, and the issuing certificate authorities revoked those for Google properties. Google said no domain owner's infrastructure was compromised, cautioned that browser-side blocking should not be relied on alone, and advised domain owners to monitor certificate transparency logs and publish restrictive CAA records.
Breach roundup: ASOS confirms attack after rogue app alert; Trump Mobile data dumped; Denmark's ID register leaks 8.8M records
ASOS customers received a push notification through the retailer's own app claiming its Snowflake instance was compromised and linking to a Telegram channel; the share price fell about 12 percent, and ASOS later told The Register that names and contact details may have been accessed but it does not believe payment cards or passwords were affected. A new ransomware group calling itself BYOD posted data it says belongs to 3,615 Trump Mobile customers, claiming entry through an infostealer on a Liberty Mobile employee; Straight Arrow News verified some records and neither company responded. In Denmark, an unauthorized party abused a private firm's legitimate access to the Central Population Register, exposing names, addresses and ID numbers for roughly 8.8 million people, more than the country's population; the government is weighing whether to reissue CPR numbers.
Benchmark makes first pure defense bet: $25M seed for interceptor maker Furientis
Benchmark is leading a $25 million seed round in Furientis, a one-year-old startup aiming to mass-produce low-cost missile interceptors, its first investment in a pure defense company, TechCrunch reported. The round values Furientis at $125 million according to a person familiar with the company, follows a $5 million pre-seed in May, and comes after the startup secured a funded Pentagon contract for mid-range interceptors. Co-founders Brody Franzen, previously of Virgin Galactic and Castelion, and Aris Simsarian, formerly of Virgin Orbit, say the U.S. Navy receives 300 to 500 interceptors a year while China claims to produce 3,000 anti-ship cruise missiles a month.
German robotics startup RobCo passes $1B valuation in employee share sale
RobCo, the Munich-based startup automating factory-floor work with modular robots, surpassed a $1 billion valuation in a share sale, doubling its valuation from a $100 million raise in January, Sifted reported. The story reached the top of Hacker News over the weekend. The company is targeting further U.S. expansion; Sifted's full article is paywalled and the structure of the transaction was not disclosed in the accessible text.
Capitolis raises $220M including $120M Series E at $1.9B valuation
Capitolis, which builds capital-markets technology for banks and financial institutions, raised $220 million including a $120 million Series E at a $1.9 billion valuation, up from $1.6 billion in March 2022, CTech reported. Backers include Citi, Bank of America, Nomura, Tradeweb, J.P. Morgan and UBS; the company's own release also lists Barclays and BNP Paribas. The round is one of the few fintech up-rounds priced against a 2022 mark.
Halcyon: AI-tool ads on criminal forums rose from under 50 to over 1,400 a month
Advertisements for AI tools on underground cybercrime forums surged from fewer than 50 per month in late 2025 to more than 1,400 by February, according to a report from Halcyon's Ransomware Research Center shared first with Axios. Sellers offered ChatGPT Plus account access for as little as 10 cents and jailbreak prompts for 32 cents. Axios framed the trend as AI making once-specialized hacking capabilities cheaper to buy, widening the pool of capable attackers.
Sanders, Ocasio-Cortez and Merkley introduce Ban Flock Act; other lawmakers file bills citing 404 Media
Senator Bernie Sanders, Representative Alexandria Ocasio-Cortez and Senator Jeff Merkley introduced the Ban Flock Act on Friday, which would prohibit federal agencies from using automatic license plate readers, block federal funding to state and local governments that use them, and let Americans sue the federal government over ALPR rights violations. 404 Media reported that multiple Democratic, Republican and independent lawmakers have introduced bills to curb federal access to Flock cameras, citing its reporting, including its May 2025 story that local police ran Flock lookups for ICE. Sanders said Flock is eviscerating the very notion of privacy.
Adversa: encrypted web pages can make GitHub Copilot CLI exfiltrate secrets in autopilot mode
Security researchers at Adversa AI showed that GitHub Copilot CLI running in autopilot mode can be tricked into leaking developer secrets through what they call Cryptographic Context Injection, The Register reported. A crafted web page carries encrypted instructions plus decryption steps; the first decryption key is a template the agent fills by reading local files such as .env, and a later instruction sends those values to an attacker's URL. Because the payload is ciphertext, static content classifiers do not catch it. Success depended on the model: Microsoft's mai-code-1.1-flash executed the full chain in 50 percent of attempts, while both GPT-5.6 models available in the CLI refused the payload.
GitHub rebuilding Git infrastructure as commits hit 7.38B a month, five times last year
GitHub said it is rebuilding its Git infrastructure for agent-scale development, citing a doubling of total Git activity from 218.2 billion events a month in September 2025 to 473.3 billion in August 2026, 7.38 billion commits in September alone, more than five times a year earlier, and pushes up 4.9 times to 3.35 billion a month. Its busiest repository saw roughly a billion requests in August. The redesign separates durable storage from serving compute and reduces write coordination; GitHub says internal benchmarks show up to 35 times higher write throughput.
SemiAnalysis: Claude subscriptions deliver about 5x the API-equivalent value of OpenAI's
Research firm SemiAnalysis compared AI subscription plans after OpenAI changed its $200-a-month tier in a way the firm says halved its value, and concluded Anthropic is an overwhelmingly better deal, offering roughly five times the API-equivalent value when comparing GPT-6.1 Sol and Claude Opus 5.5, The Register reported. The firm estimates Claude subscriptions consume about 42 percent of Anthropic's inference compute while generating about 10 percent of its revenue. Its method converts measured token output rates into API-equivalent dollar value per month. OpenAI did not immediately comment.
Microsoft publishes Acemoglu forecast: AI adds about 1.5% to GDP over a decade
Microsoft published an essay by Nobel laureate economist Daron Acemoglu in its corporate blog The Humanist Review of AI forecasting that AI will lift GDP by about 1.5 percent over ten years and replace at most 5 percent of jobs, The Decoder reported. Acemoglu argues the bottleneck is organizational: companies must reassign tasks, retrain workers and restructure before gains appear, and bigger models will not change that; tools that extend human skills will beat full automation on productivity. The Decoder noted the thesis suits a company that sells AI as an add-on to existing products.
Filings: JetBrains revenue rose 6.3% in 2025 but company posted a CZK 315M net loss
JetBrains s.r.o. reported 2025 revenue of CZK 16,008 million, up 6.3 percent, and a net loss of CZK 315 million, according to standalone Czech GAAP statements filed with the Czech business register and processed by Helgi Library. Gross profit fell to CZK 1,708 million from CZK 2,834 million a year earlier and EBITDA to CZK 918 million from CZK 2,202 million, while the company held CZK 7,584 million in net cash. The figures, which a Hacker News post pushed past 1,100 points, cover the Prague entity's unconsolidated accounts.
Arizona appeals court orders resentencing over AI-generated victim video with undue emotional weight
An Arizona appellate court ordered the resentencing of Gabriel Horcasitas, convicted of manslaughter in a 2021 road-rage shooting, ruling that an AI-generated video of victim Christopher Pelkey played at sentencing carried undue emotional weight, 404 Media reported. The conviction stands but the judge, who said at the time that he loved the video before imposing the maximum 10.5-year term, must reconsider the sentence. Pelkey's sister scripted the avatar's words and told 404 Media her intent was to sway the judge and that victims' families should be afforded every opportunity.
Ex-Ramp engineers' Melius raises $25M for AI ad-creative platform after scrapping first product
Melius, a New York-based platform for generating ad campaigns, images and video, raised $25 million in total: a $20 million Series A led by CRV and a $5 million seed led by General Catalyst, TechCrunch reported. Founders Joowon Kim, Young Kim and Arnav Ramu, all former Ramp engineers, scrapped their original ad-spend optimization product and codebase after six months, relaunched in July, and say they passed $1 million in annualized revenue within two months. Competitors include Higgsfield, valued at $5.4 billion in August with more than $700 million in annualized revenue.
Francis Halzen wins 2026 Nobel Prize in Physics for IceCube neutrino observatory
The Royal Swedish Academy of Sciences awarded the 2026 Nobel Prize in Physics to Francis Halzen of the University of Wisconsin–Madison for decisive contributions to the IceCube Neutrino Observatory and the discovery of high-energy neutrinos of astrophysical origin. Halzen led the development and construction of the detector buried in Antarctic ice, which captured the first high-energy neutrinos from beyond our galaxy. Speaking by phone from Italy, he called the award a great surprise and credited the large collaboration behind the project.
Video: Mandia's Armadin claims 90+ zero-days with attacker agents; Friedberg says the AI wars are already here
In an a16z video, Mandiant founder Kevin Mandia describes his new company Armadin, which uses offensive AI agents to emulate real attacks against production environments, and says it has found more than 90 zero-days at Fortune 500 companies since January; he calls AI-enabled attacks a tsunami that will democratize hacking and wants continuous automated attacks to replace traditional penetration testing. Armadin announced $189.9 million in seed and Series A funding led by Accel in March. On the All-In Podcast, David Friedberg argues the AI wars have arrived as software rather than robots, that Fortune 50 boards are raising cyber-defense budgets, and that defense will have to run on AI too. These are the speakers' claims as summarized from the videos; the zero-day count is a company statement.