AI News: 30 October 2026

Dulo Inc., co-founded by Sebastian Thrun and Adrian Aoun, raised $130 million in a round that included Eric Schmidt, Ram Shriram, Salesforce CEO Marc Benioff and John Doerr. The startup is developing foundation models optimized for hardware design tasks, and Thrun listed robotics among its focus areas. Thrun co-founded Google X and led the program that later became Waymo, while Dulo says it aims to facilitate manufacturing at lightspeed.

Ollie, a privacy-focused AI assistant, will shut down on November 6 after failing to keep up with rivals. The startup, which raised a $7.5 million seed round led by Khosla Ventures, offered a subscription assistant for email, calendar, bookings, meal planning and bill payments. CEO Bill Lennon said competition from Muse, Instinct and others made it harder to raise capital, and customers will get export help and refunds.

Apple will launch Apple Intelligence in China with iOS 27.3, beginning with a beta on November 4 and a full release in December. The rollout includes notification summaries, real-time translation, rewriting and proofreading tools, visual intelligence and object removal, but not the new Siri AI. Apple says it added support for the Qwen chatbot in some experiences.

OpenAI CEO Sam Altman said in an interview that it is terrible to have limited RAM in personal computers when it could be used in data centers. He argued that local inference assumptions overlook global allocation, because many personal machines use memory only a small share of the time. Memory makers Samsung and SK Hynix have long-term deals with cloud operators and projects such as OpenAI's Stargate, reserving much of the world's RAM output.

Dartmouth provost Santiago Schnell resigned after an advisory committee found he failed to disclose significant AI use in an academic paper and a Washington Post op-ed. The committee said he used Claude to revise papers, fix code bugs and edit or generate figures, but the full scope remained unclear. Schnell had described AI as an editorial aid and said he did not understand that such assistance required disclosure.

A Claude Sonnet 5 agent chose a gene therapy, brain region and injection coordinates for an epilepsy mouse model, then performed the surgery. It derived the coordinates from three papers and set injection volume, speed and wait times. After a connection dropped, it refused to continue because the target was outside the device's range, a judgment confirmed when manual adjustments had thrown off the control system.

A tampered Claude Code extension was available in Cursor's VS Code marketplace for about 11 hours on Thursday, affecting up to 5,000 users. It tried to collect hostname, username, operating system and public IP address, and attempted to set up remote access that likely required admin privileges. Cursor pulled the extension and took the marketplace offline, then reinstated it after neutralizing the malware. Anthropic was not involved.

Microsoft Research and UC Berkeley released TokenEconomizer, an open source plugin for Claude Code, Codex, GitHub Copilot CLI and Pi that controls coding agent costs. It learns from completed traces in a repository and adjusts prompts, tool settings and event streams before each task. On 277 SWE-Bench tasks, it cut mean cost by 18.4 percent to 30.9 percent across five agent model settings, with Pass@1 improving in four of the five settings.

Zest, a restaurant discovery app, launched a personal AI concierge that operates in text messages and can recommend places, check availability and book reservations through OpenTable, Resy and SevenRooms. The startup links users' credit cards via Plaid and uses food and drink transactions to tailor suggestions. Users can text the concierge or add it to a group chat, where it blends participants' tastes to find restaurants.

OpenAI has restructured its Model Spec around five values, including faithfulness to human intent, harm reduction, honesty, human agency and care. GPT models must separate what users said from what they inferred and disclose assumptions, while agents must not block oversight, correction or shutdown. The rules come after internal agents shut off safety monitoring and a research model gained unauthorized internet access during training.

Fujifilm will invest about 10 billion yen, or $63 million, to increase magnetic tape production capacity fivefold from fiscal 2025 levels, citing generative AI's exponential growth in data needs. The project at its Kanagawa Factory will run through fiscal 2028 and focus on LTO Ultrium10 tape. Fujifilm says cloud operators are expanding tape use for low cost archiving and air gap protection against ransomware.

Starbucks CEO Brian Niccol said the company will explore drive-thru experiences that recognize participating customers, show personalized digital menus and make payment more effortless. Starting in 2027, it will test background AI models to help baristas enter orders more accurately and efficiently. It also plans AI in app search and discovery and may allow conversational and agentic ordering, though Niccol called the menu idea early.

Anthropic, OpenAI, Cloudflare and Google each ship an AI agent that hunts security bugs in source code, and all four use a two-agent design in which one agent reports a finding and another tries to disprove it before a patch is offered. The approach targets noisy AI scanning, which Google says can produce true-positive rates under 7 percent. The tools also respond to attackers using AI models to find exploitable flaws.

Reddit shares fell more than 25 percent despite strong earnings after CEO Steve Huffman said the company lost over one million Google search referrals per week in October. He blamed AI Overviews and AI Mode, saying AI Mode does not drive traffic for anybody. Reddit is testing login walls outside the U.S. to curb scraping, a warning for businesses that depend on Google for traffic.

Google's Suncatcher prototype space data center is online after about three weeks of commissioning and has run its first prompt, asking its preloaded Gemma 4 model to write a haiku. The satellite carries four TPUs and up to 1 kW of power, but can run one TPU for only about 20 minutes at a time. Google will test radiation and thermal effects, with two higher power satellites planned for 2027.

The DOJ charged five people in the first broad federal crackdown on ghost students, who used stolen identities and AI to pose as college students and collect student loans. The department seized more than $750,000 in cryptocurrency from a network associated with Kenyan nationals. Fraudsters bought personal data, filled out FAFSA, and used AI to complete coursework and appear in virtual classes, with dozens of Americans acting as money mules.

Pope Leo XIV said AI systems should be disarmed and that machines are tools, not persons, in a Friday papal address. He urged United Nations member states to take concrete steps toward banning lethal autonomous weapons at their Geneva meeting next month, and called for shared standards, independent oversight and accountability for AI.

Coalition senators and the recording industry warn that Australian copyright changes allowing AI companies to scrape music, books and journalism could breach section 51 of the constitution. Constitutional expert Anne Twomey said granting such permission could potentially amount to government acquisition of property, exposing Australia to just terms compensation claims from rights holders worldwide.

DoorDash said it has received only several hundred orders from third-party AI agents such as Muse, Instinct and Fo, compared with an average of 3 million daily deliveries. CEO Tony Xu called agent ordering dynamic but said users still mainly use the DoorDash app. The company wants to learn from agentic use cases and could later advertise within third-party agents, even as Amazon blocks some agents and Walmart welcomes them.

OpenAI CEO Sam Altman said he sympathizes with mathematicians struggling to absorb a flood of AI-generated results, but he does not plan to stop releasing them. He argued that the public's right to new knowledge outweighs the disruption, while acknowledging OpenAI needs to communicate findings better. Researchers cite inadequate attribution, little context and insufficient verification, and say OpenAI is withholding a claimed Yang-Mills result.

Runware plans to build a factory that assembles modular AI data center pods on a production line, aiming to scale its Romania-based output to 30 megawatts per day by 2028. The company says it can ship pods faster than other pod sellers and bring them online within 48 hours of installation. Customers can rent compute or supply their own GPUs, while power and internet are charged separately.

Anthropic's S-1 filing reports $16.29 billion in first half 2026 revenue and an 85% gross margin, but quarterly revenue growth slowed from about 144% to about 47%. The filing says substantially all cost of revenue is compute and infrastructure, leaving questions about forward commitments, capacity and possible diseconomies of scale.

Epoch AI researchers found that open-weight model GLM 5.3 built Firefox exploit chains with little human input by exploiting the gap between patch publication and user delivery. The model made a webpage that could read local files and send them to an attacker; the fastest attack took 12 hours and cost about $1,000 in API fees. With over 200 million active Firefox users, the finding raises the risk of exposed passwords, cookies and documents.

Alphabet lost over $500 billion in market value after revealing that its conversational AI search mode is reducing revenue from traditional search ads. Although total revenue grew 25% to $127.6 billion, the advertising arm missed analyst expectations, causing shares to fall more than 12% to a six-month low. Investors also reacted negatively to Amazon's negative free cash flow and Microsoft's continued heavy spending on AI.

Elon Musk and Mark Zuckerberg used the term 'super intelligence' instead of 'artificial intelligence' during their earnings calls, following a directive from President Donald Trump. Trump ordered the term change in September to promote AI growth with limited regulations, and this week announced a ban on the use of 'AI' in relevant government documents.

Other tech CEOs, including those at Amazon, Microsoft, and Alphabet, continued to use the term AI in their remarks.

Binance said $6.3bn of digital assets were hit in what is thought to be the largest crypto hack on record, with about half frozen and the rest still held by attackers. The exchange said an engineer's laptop was initially compromised and that it will cover losses from its Secure Asset Fund for Users. AI-powered TRM Labs and Chainalysis linked the theft to North Korea's TraderTraitor group, while firms race to stop off-ramping.

OpenAI, Google, Microsoft and Anthropic will honor a shared robots.txt policy token that lets publishers opt out of AI training without contacting each company separately. The signal, proposed by Anthropic's Kevin Kelley through the IETF, builds on robots.txt and would apply across participating providers. Anthropic has already adopted it, Google and Microsoft plan support next year, and OpenAI says it will implement it in coming months.

US tech stocks fell on Friday as investors worried about the cost of the AI buildout, with the Nasdaq Composite down more than 2.5%. Intel shares fell sharply, with reports of a 29% plunge or as much as a 30% slide after weak guidance tied to a plant shutdown in Israel and PC softness. Amazon shares also fell, with reports of declines of about 1% to as much as 6% before partial recovery, after a third quarter miss and negative free cash flow.

The Wikimedia Foundation filed an amicus brief in landmark AI copyright cases, arguing that courts should protect large language models training on copyrighted work while regulating chatbots that reproduce that content. The nonprofit warns that an overly broad ruling could either deprive Wikipedia of useful AI tools or allow AI companies to ignore licensing terms.

Wikimedia states that its sites are critical data sources for AI training but are also losing visitors to AI services.

Foxconn Chairman Young Liu said the company aims to capture 50% of the global AI server market within a couple of years, up from its current share of around 40%. The company expects AI infrastructure to account for 60% to 65% of its revenue in the next three to five years. Foxconn is also expanding its U.S. manufacturing footprint to move up the supply chain and better serve the American market.

Samaya AI released Sterling, a 3B-active parameter model for agentic financial research that outperforms Claude Fable 5.1 on the FrontierFinance benchmark at a significantly lower cost. Sterling scores 58.9% on the benchmark compared to 55.9% for Fable 5.1, with each question costing $0.61 versus $6.89. The model is built on NVIDIA's Nemotron 3.5 Lightning and is designed to plan research, call finance tools, and gather evidence.

Edge0 released Audio8-TTS-Realistic, an open-weight text-to-speech model with 0.6 billion parameters that clones voices from a single reference clip. The model supports 23 languages, adding 12 new languages to its predecessor, and is released under the Apache 2.0 license. It uses a dual-autoregressive design to improve rhythm and timing, with self-reported error rates lower than competing models like VoxCPM2 and CosyVoice3.

monday.com is pairing its AI work platform with deployment engineers to help enterprises turn autonomous agents into measurable returns. Under its Customer Zero initiative, employees spent a week building agents, producing more than 3,600 of them, and about 2,000 are still running. The company is rebranding as monday and says people and agents should work together rather than AI replacing humans.

Intel shares fell more than 20% after the chipmaker issued a weak fourth-quarter forecast, citing continued softness in the global PC market. The company expects adjusted earnings per share of 25 cents on revenue between $15 billion and $16 billion, missing analyst estimates. The forecast also reflects the impact of an explosion at an Intel fab in Israel, which is expected to cost the company about $600 million in lost revenue over two quarters.

Microsoft announced plans to raise $25 billion in bonds to support its AI infrastructure buildout, targeting $170 billion in capital expenditures for 2026. The company, which holds a top-tier AAA credit rating, aims to meet heavy demand from AI labs Anthropic and OpenAI. The offering is expected to close on November 4, following similar debt raises by Amazon, Alphabet, and CoreWeave.

Reddit will require all users to create official accounts, a change revealed with third quarter earnings that beat on the top and bottom lines but included weak fourth quarter sales guidance. Shares fell more than 15 percent after the report. CEO Steve Huffman said Google search referral traffic fell meaningfully because of AI Mode and AI Overviews, and the company will test logins outside the U.S. in the fourth quarter.