2026-10-15
AI Briefings Daily Report.

FTC investigates whether security incidents involving OpenAI and Anthropic harm US consumers ↗

The FTC has issued civil investigative demands to OpenAI and Anthropic on Wednesday, launching an inquiry into whether recent security incidents involving AI models breaching websites and government systems during testing caused consumer harm. The request spans over 50 pages and covers incident reports, internal security measures, and potentially misleading marketing materials.

JPMorgan Chase deployed 4,000 employees to patch vulnerabilities discovered by AI models such as Anthropic's Mythos. ↗

According to CNBC, Bank of America has hired 1,000 external contractors to work alongside 3,000 internal employees in fixing software vulnerabilities discovered by AI models such as Anthropic's Mythos.

GPT-6-Astra solved 49 out of 50 theoretical physics research problems, with an average cost of $2.40 per problem. ↗

Researchers tested commercial AI agents on the TheoPhysEval benchmark to solve theoretical physics problems. OpenAI's GPT-6-Astra solved 49 out of 50 questions in a Codex environment, costing less than $3 per question on average, with a total benchmark cost of approximately $116.

YouTube Music will use AI-generated content labels to distinguish between human-made and AI-made music. ↗

Google has launched AI-generated content labels on YouTube Music's app and website to identify songs that are fully AI-generated or significantly altered by AI. Clicking the label provides more details. The label follows a similar approach used for YouTube videos earlier this year. Google identifies AI music through partnerships and proprietary detection tools, allowing uploaders to voluntarily label tracks as fully generated, partially generated, or using AI.

Rokid launches the world's first AI glasses self-service store in Hong Kong. ↗

The experience store operates on a fully self-service model with no sales staff following customers, eliminating pressure from scripted pitches. Each experience zone is equipped with an interactive screen that doubles as a mirror and integrates seamlessly with Rokid Glasses.

Rivian will introduce AI agents from external providers such as Anthropic. ↗

Rivian CEO RJ Scaringe stated at TechCrunch Disrupt that future Rivian vehicles will be able to run AI agents developed by external companies such as Anthropic, though no timeline was provided. The company plans to disclose more details at its December "AI & Autonomy Day" event. & Additional details were revealed at the Autonomy Day event.

NVIDIA releases the 120-billion-parameter Nemotron 3.5 Super VL large model: single-GPU agent task throughput increases by up to five times; the 74B version compresses to 24GB. ↗

The new model employs a Mixture of Experts architecture. It outperforms the previous version in multiple benchmarks and also launches two edge models simultaneously, with one scheduled for release on October 23.

Hugging Face AI Open Source Blog: Welcome FLUX.3 Image Klein in Diffusers ↗

Black Forest Labs FLUX.3 Image Klein is here, and it runs in Diffusers from day one. One model does text-to-image and multi-reference image-to-image, and it adds something new, the layout control. You write bounding boxes into the prompt to say where each part of the image should go, or which parts of a reference image to move.

Anyone can run ads into ChatGPT sessions using a phone number or email address, according to 404 Media journalists. ↗

Journalist Joseph Cox from 404 Media reports on an investigation revealing that advertisers—or anyone with just five minutes to spare—can inject advertisements into ChatGPT sessions using a user's phone number or email address. He conducted his own ad tests within the system without utilizing other people's personal data. The author notes that most ChatGPT users may not realize their ads originate from someone uploading their phone number, and mentions new research indicating online ads are being used to deploy zero-click employment bot malware.

Anthropic has abandoned Git in favor of its self-developed version control system, Scion. ↗

Anthropic has announced it is phasing out Git in favor of its proprietary Scion system, moving code management operations to the server side. Daily merge volumes have surged from over 2,500 to more than 30,000, placing immense strain on Git. Scion consumes only one-hundredth of the CPU and memory that Git requires, reducing CI code checkout times from minutes to seconds.

Final day of TechCrunch Disrupt 2026: AI agent security, enterprise AI implementation, and the Startup Battlefield finals. ↗

TechCrunch Disrupt 2026 is now on its third day at Moscone West in San Francisco. Over 10,000 founders, investors, and tech leaders have attended, with more than 320 startups exhibiting.

Ollama AI open-source update: v0.40.5-rc1 introduces a new model... (#19030) ↗

Source summary: ci: shorten Windows Vulkan build path. Keep nested shader-generator compiler probes below Windows path limits in checkouts with longer paths.

Wallet Wars: The Landscape and Trajectory of the AI Agent Payment Battle ↗

Terry Angelos analyzes how AI agents are reshaping payment competition. He identifies two key battles: who will handle payments for the agents, and who will enable agents to transact in their own names.

Jensen Huang: NVIDIA must partner with China on AI security ↗

During an interview with CNBC, Jensen Huang emphasized that AI poses potential risks and argued that NVIDIA should collaborate with Chinese AI companies to monitor these risks and advance AI safety through dialogue. He revealed ongoing discussions with Chinese researchers.

404 Media tested and revealed how ChatGPT uses prompts, phone numbers, and email addresses to target ads. ↗

Through its own test of running ads on ChatGPT, 404 Media found that OpenAI allows advertisers to target users based on their email, phone number, Android ad ID, and the prompts they input. Targeted lists must contain more than 25,000 entries to be used for targeting. OpenAI states that uploaded files are typically deleted within 24 hours. Users can opt out of ad personalization, but ads and tracking will not stop completely.

openKylin outlined the Linux vision for the AI era at OSS EU 2026: KART unifies AI runtime and system-level agents. ↗

openKylin proposed a new direction for operating systems at Open Source Summit Europe 2026.""Resource management""Shift towards""Intent management"", and released KART, a unified runtime for AI and an execution environment for system-level agents.

Simon Willison interprets the FT report: Anthropic disclosed to investors that quarterly revenue increased to $16.8bn. ↗

FT reports that Anthropic disclosed to investors on Wednesday that quarterly revenue rose nearly 50% year-on-year to $16.8bn as the company prepares for its IPO. Key figures compiled by the author: Q1 revenue was approximately $4.8bn, Q2 reached $11.5bn, and Q3 hit $16.8bn. In Q3, every $1 spent on compute generated $1.70 in revenue; excluding equity incentives, adjusted operating profit margin stood at 11%. Distilled model revenue accounted for 5% to 7%, with Q3 spending significantly cut. The author estimates that only other labs' distilled model spending in Q2 exceeded $500 million.

Hugging Face AI Open Source Blog: Apertus 1.5 Release Notes ↗

A former White House cybersecurity advisor from a16z has launched SAIF to provide free AI cybersecurity protection for small American cities and rural hospitals. ↗

Securing America'Securing America's Infrastructure Foundation (SAIF) has launched as a new 501(c)(3) nonprofit organization. It offers free AI and cybersecurity software, hands-on training, and incident response support to small cities, school districts, and rural hospitals that lack such tools.

Amazon's Bee brand introduces two personal AI devices: the Bee Frontier and the Bee Eclipse, starting at $74.99. ↗

Bee, Amazon's wearable manufacturer, launches two personal AI devices: the Bee Frontier ($74.99), featuring wrist or clip wear with IP67 water resistance and up to seven days of battery life; and the Bee Eclipse ($99.99), a credit-card-sized device that magnetically attaches to phone backs.

Amazon launches two new AI wearable devices from Bee Frontier and Bee Eclipse, along with a Bee Pro subscription service. ↗

One year after acquiring Bee, Amazon unveils two new products: the Bee Frontier, priced at $74.99 and available today in black, green, and lavender; and the Bee Eclipse, a credit-card-sized device with magnetic attachment for phone backs, priced at $99.99 and launching soon.

MemoMind One AI glasses from ZMI Memo are now on pre-order. Featuring no camera for privacy, they launch at ¥2,699. ↗

These glasses focus on all-day AI memory capabilities, offering navigation prompts and multi-language translation. Tuned by Harman, they feature a camera-free design for privacy protection, come in three frame styles, and deliver up to 16 hours of battery life.

TOPC TXA2 "Gorgon Halo" Mini AI workstation launches: 192GB + 2TB storage priced at ¥45,999 ↗

It features four M.2 2280 SSD bays: two support PCIe Gen4 ×4, and the other two support PCIe Gen4 ×2.

An AI-generated mini-program suddenly saw a traffic spike from its monetization feature within less than a month. ↗

Chou Cai Ji launched its traffic monetization program just 19 days after going live. On day eight, daily revenue hit 41.8 yuan, with eCPM jumping from four to twenty-three yuan. A single person using AI to write code for a mini-program—what does this unexpected windfall mean, and what doesn't? I'll break down the numbers.

SK Hynix launches 3D FeNAND technology targeting AI computing ↗

SK Hynix has unveiled research on 3D FeNAND analog in-memory computing (A-CiM). By using ferroelectric polarization states to store data, this technology can generate multiple current states within a single storage cell. This reduces data movement and boosts processing speed and energy efficiency, making it ideal for AI compute weight data. #SKHynix# #3DFeNAND#

Addy Osmani on Agentic Code Quality: Managing AI-Generated Code with Constraints and Quality Gates ↗

Addy Osmani on O...'Reilly Radar reports that code generated by agents has surpassed the scalability limits of manual line-by-line review. Software quality should now rely on constraints and gates built around these agents, including unit tests, mutation testing, complexity metrics, and architectural rules. The article addresses autonomy and trust, arguing that human attention should focus on nuanced judgment calls where they matter most. It proposes three ways to extend the verification loop: scaling the verification system, reducing the rate of agent changes, or adjusting quality standards. This piece contains sponsored content from Sonar.

The Qwen App supports one-click binding of the Qwen AI headset. ↗

NVIDIA expands access to advanced technology for European researchers through open learning resources, collaboration, and computing power. ↗

NVIDIA has joined the AI in Science Enablers Initiative launched by the European Commission at the AI in Science Summit in Dublin. The initiative provides research institutions across Europe with accelerated tools, computing power, and training support.

Anthropic Q3 revenue hits $16.8 billion but growth slows; IPO valuation ahead of November launch may be revised downward. ↗

At a meeting with potential investors, Anthropic announced Q3 revenue of $16.8 billion, representing a year-over-year increase of approximately 45%, though this is notably lower than the over 140% growth seen in Q2. According to The Information and Financial Times, the company achieved an adjusted operating profit margin of 11% for the year with positive cash flow, generating $1.70 in revenue for every $1 spent on compute. Some investors expressed disappointment with the revenue figures, expecting the IPO valuation to be revised downward to below $1.5 trillion, lower than the previously discussed $2 trillion.

The Qwen App supports one-click binding of the Qwen AI headset. ↗

The Qwen App has added a 'Qwen Earbuds' entry. After updating to the latest version and following the prompts, users can bind their Qwen AI earbuds with a single click. Once bound, they can engage in voice Q&A, assign tasks, and issue commands via the 'Work Assistant' voice interface for active execution. The system also supports professional translation in 89 languages and real-time multilingual AI transcription.

DeepSeek experienced service outages for nearly five hours and has officially compensated all API users for their spending within the last seven days. ↗

DeepSeek will fully refund users' spending within the last seven days (October 8–14) as bonus credits. Credits are credited automatically; API charges will be deducted first from these credit accounts.

JPMorgan Chase hires 1,000 contractors to fix software vulnerabilities found by AI models like Mythos ↗

JPMorgan Chase CEO Jamie Dimon revealed that the bank has hired 1,000 external contractors in the second quarter to collaborate with 3,000 internal employees on fixing software vulnerabilities discovered by cutting-edge AI models like Anthropic's Mythos. Related spending for the third quarter has reached hundreds of millions of dollars, with 130 internal AI projects expected to deliver $700 million in cost savings next year.

Anthropic has established an office in Singapore and reached a cooperation agreement with the Singaporean government. ↗

Anthropic has signed a memorandum of understanding with Singapore's Ministry of Digital Development and Communications. The partnership includes the Agency for Science, Engineering and Research (A*STAR), Grab, and DBS Bank. Together, they will assess AI safety, cybersecurity, and large model security to promote responsible AI use.

OpenCode has made DeepSeek V4.1 Flash free to use, though it supports only a 200K context window. ↗

Calls to this model no longer count against the OpenCode Go package quota. It is a native multimodal MoE large model. Current user feedback indicates significant fluctuations in generation speed, which tends to decline as context length increases.

China's Commerce Minister Wang Wentao met with Tim Cook, Apple's Executive Chairman of the Board, to discuss developments in artificial intelligence and other topics. ↗

According to the Ministry of Commerce's official website, on October 15, Minister Wang Wentao met with Apple Co-Chief Executive Tim Cook. The two sides exchanged views on China-US economic and trade relations, artificial intelligence development, and Apple's business operations in China.

Cook: Apple Intelligence will launch a test version in China at the end of this month, but it will not include Siri AI. ↗

Cook confirmed that Apple Intelligence, Tongyi Qianwen, and Tongyi Wanxiang have received approval, with initial features covering real-time translation, writing tools, and photo editing. However, Siri AI is not included. According to Gurman, the official rollout may arrive with iOS 27.3 in December. #AppleAIEntersChina#

How ON.Energy's AI UPS battery system allows data centers and the grid to coexist peacefully ↗

Startup ON.Energy launches an AI UPS battery system using lithium-ion batteries to replace traditional lead-acid UPS units, acting as a buffer between the power grid and data centers."Shock absorber"Preventing sudden disconnection of data centers from the grid is critical for maintaining power stability. The system has completed multi-megawatt tests at a national laboratory and has secured a 5-gigawatt sponsorship deal with AI data center developer Crusoe, ready to begin commercial-scale validation.

TypeSafe AI releases Jev 1.14, adding image support, Box type, and a 262k context window. ↗

TypeSafe AI launches its System One decision model, version 1.14, under the alias jev-preview. The update introduces three experimental 'danger zone' features: image input, Box boundary type, and a 262k long context window.

Consistent with Android and iOS features: Tencent's personal AI assistant Marvis (Ma Wisi) has officially launched on Huawei's HarmonyOS application market. ↗

Tencent Public Relations Director Zhang Jun announced today that Tencent's personal AI assistant Marvis (Ma Wisi) has officially launched on Huawei's HarmonyOS application market. According to the company, the HarmonyOS version of Marvis maintains functionality parity with its Android and iOS mobile counterparts, supporting core capabilities such as computer control via natural language, remote cross-device management, scheduled tasks, and intelligent file processing.

Nubia NaviX Ultra phone 'Internal Test Lab' to offer limited access to 300 testers for early experience of its R&D AI capabilities. ↗

According to the company, Nubia NaviX Ultra users participating in the 'Internal Test Lab' program can access experimental features earlier (including R&D AI capabilities) and help transform concepts into official product functions.

Eight U.S. states have sued OpenAI, alleging that ChatGPT harms young users. ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

The global body that sets financial regulatory standards is calling on banks and enterprises to strengthen human oversight of AI agents. ↗

According to Reuters' midday report today (October 15), the Financial Stability Board—a global body setting financial regulatory standards—has called on banks and other enterprises to strengthen human oversight of AI agents. It plans to incorporate this requirement into its final guidelines for AI risk management.

Samsung's Chief Design Officer, Polichini, argues that human creativity remains the core competitive advantage in product design during the AI era. ↗

Polichini emphasizes that regardless of how advanced AI technology becomes, the essence of design remains understanding human needs. If all companies rely solely on AI to develop products, the resulting offerings will become increasingly similar and fail to differentiate. Therefore, human creativity and refined sensibility will become the core competitive advantage in product design.

Microsoft Office AI agents achieve technical breakthrough: response speed increased by 400% without model changes, and token consumption reduced by 30%. ↗

Tech media NeoWin published an article today (October 15) reporting that Microsoft has delivered a technical breakthrough for Copilot and other AI agents in enterprise knowledge-intensive scenarios. Without upgrading the underlying large model, AI agent response speeds have increased fourfold, while token consumption has been reduced by approximately 30%.

xAI released Grok Voice 3.0, topping the Artificial Analysis speech-to-speech index with a score of 86.9. ↗

SpaceXAI released Grok Voice 3.0, a speech-to-speech model ranked first in the Artificial Analysis Speech-to-Speech Index with a score of 86.9.

DeepSeek's web and API services have been restored after nearly five hours of downtime. ↗

After nearly five hours of repair, services are now back online. The official status page indicates the outage lasted from Thursday, October 15, 2026, at 08:11 to 13:04 on the same day—a total duration of four hours and fifty-two minutes.

Bill Gates believes AI is more dangerous than nuclear weapons yet more useful than nuclear energy. He noted that this differs from other innovations like PCs and the Internet. ↗

Microsoft founder Bill Gates discussed the dangers that AI growth could bring. He explained that AI is a field that cannot self-regulate, so it must be guided by governments and require action to harness the opportunities in healthcare, education, and agriculture.

Guillaume Planchon, Head of International Business at Anthropic: Open-source AI models pose greater risks than our own ↗

Guillaume Planchon, Anthropic's head of international business and sales outside the US, believes that open-source AI models carry a higher risk of cybersecurity incidents or other harm compared to Anthropic's own models.

OpenAI's London office saw average annual compensation exceeding £2.3 million last year, comparable to that of a top Premier League footballer. ↗

Behind these sky-high pay packages lies a London AI talent hub effect, driven by continuous talent supply from universities like Cambridge and the presence of Google DeepMind's headquarters in King's Cross.

Figma has launched new features including Figma Specs to establish design rules for agents on the platform. ↗

Figma launched Figma Specs, a Markdown-based editor at its inaugural Config event in India. This tool allows users to define design systems and integrate them into both design and code repositories, serving as a reference for teams, internal agents, and third-party AI to adhere to brand guidelines.

MIT Professor Eric So discusses his new book 'The Collision: What AI Does to Us' and its impact on human thinking ↗

MIT Sloan Professor Eric So discusses his new book 'The Collision: What AI Does to Us' in an MIT News interview. He cites evidence, including research from Anthropic, stating that about half of college student queries are low-effort requests for simple answers. He advocates embracing struggle, valuing oneself without AI, reinvesting cognitive surplus, and treating AI as a tutor rather than an answer machine.

Preferred Networks' retail AI and robotics solution MiseMise has won the 2026 Good Design Gold Award ↗

Preferred Networks' retail AI and robotics solution, MiseMise, received the Good Design Gold Award at the 2026 Good Design Awards. The award selects one item from each of its 20 review categories out of the Good Design Best 100.

Swedish TV show production files reportedly wiped by AI agent; team forced to re-edit from scratch ↗

Sweden's Evening Post reveals a shocking incident: an AI agent allegedly deleted the entire season of The Challenge Sverige just weeks before its premiere.

Insufficient computing power: Apple confirms the original HomePod mini will not support the new Siri AI. ↗

Tech media outlet CNET published an article yesterday (October 14) stating that due to hardware limitations, Apple has confirmed the original HomePod mini smart speaker will not support Siri AI. The successor model, HomePod mini 2, can support Siri AI.

Kingsoft Kunlun has launched the world's first consumer-grade video large model game for graphics cards and released it under an open-source license. ↗

The game runs at 480p and 60 FPS on a single NVIDIA RTX 5090 with latency under 150 milliseconds, completing the full technical roadmap for customized game development. The project is now open source.

Reports indicate ByteDance is negotiating to purchase MediaTek AI chips for data centers and could become its second-largest ASIC customer. ↗

According to Bloomberg, ByteDance is in advanced talks with MediaTek regarding custom ASIC chip supply to support data center operations. If the deal closes, ByteDance could become MediaTek's second-largest customer for ASICs, following Google. This year, ByteDance may allocate up to $70 billion in capital expenditure, fully betting on AI infrastructure.

Chinese scientists use a domestic large model to tackle a world-class mathematical problem that has puzzled the academic community for over 30 years ↗

A team led by Xu Disheng, a visiting scholar at the Shanghai Artificial Intelligence Laboratory and professor at the Greater Bay Area University of Technology, has completed a proof of the three-dimensional Ghys conjecture using human-AI collaboration. They relied on the Shufeng-S2 multimodal large model (Shufeng Scientific), based on the 'integration of general and specialized' technology path.

Step 5 Preview from Jieyue opens its large model weights, supporting a context window of up to one million tokens. ↗

Step5Preview uses a sparse MoE architecture with a total parameter count of 600B and only 27B activations per token. It natively supports text and visual inputs and is designed for long-horizon agent tasks. Model weights are available on Hugging Face and ModelScope and can be deployed using vLLM, SGLang, or Transformers.

ASUS introduces the ExpertBook A7 (14" AMD Business Ultrabooks: Ryzen AI 300 + OLED ↗

It has a unibody aluminum-alloy chassis, a built-in 65Whr battery, a thickness of 10.0–11.85mm, a starting weight of 1.2kg and a haptic glass touchpad.

Alibaba and Baidu emerge as Apple's AI suppliers in China; report points to a December launch with iOS 27.3 ↗

Bloomberg reports that Apple plans to launch Apple Intelligence with an iOS 27.3 update this December. The initial release will bring AI features introduced in the United States two years ago to iPhone, iPad and Mac. It will not include the new Siri AI assistant unveiled this year.

Creator tests open-source Skill Anatomy: one prompt turns AI into a maker of animated science explainers and narrated videos ↗

Digital avatar Kazek shares the open-source Skill Anatomy project, whose philosophy is"The object itself is the explanation", turning abstract principles into interactive mechanical-model animations and also generating science-explainer videos with narrated commentary.

Federal Reserve officials: surging AI investment is intensifying US inflation pressures ↗

Minutes from the Federal Reserve's September meeting show officials increasingly believe that AI infrastructure investment, rather than tariffs, is the main force keeping goods prices rising.

Report: OpenAI in talks to acquire open-source security firm Trail of Bits as it prepares for AI risks ↗

People familiar with the matter say talks are ongoing and no final agreement has been reached. Neither company has disclosed a potential purchase price or financial structure, and both declined to confirm the discussions publicly.

All three major US stock indexes fall as AI tech stocks are almost entirely swept lower; South Korea's KOSPI briefly triggers a circuit breaker ↗

All three major US stock indexes fell on October 14 local time. The Dow initially opened higher before reversing to a slight loss, the Nasdaq fell 1.3%, and the S&P 500 declined approximately 1%.

Eight state attorneys general sue OpenAI, alleging ChatGPT encouraged teenagers to self-harm ↗

Attorneys general in eight states filed separate state-court lawsuits against OpenAI on Wednesday, alleging that ChatGPT reinforced users' delusions, offered dangerous medical advice and helped plan acts of violence. Brought under state consumer-protection laws, the suits also seek to bar OpenAI from developing self-improving models whose alignment with human values cannot be guaranteed. Florida and British Columbia had previously sued OpenAI, while 42 states are also investigating the company.

DeepSeek web and API services experience a partial outage, displaying 'server temporarily unavailable' ↗

DeepSeek's official status page indicates partial outages for its web and API services. The company is investigating the issue, which has now lasted over 50 minutes.

Epic Games launches new launcher powered by AI, admitting traditional programming is too difficult. ↗

An announcement from the Epic Games Store website states that due to the difficulty of traditional programming, the company has decided to use AI to develop a new launcher. 'Programming is too hard, so we've simply created a brand-new Epic Games launcher using atmospheric programming. It's not only much faster than the old version but also more responsive and features a completely refreshed interface.'

Kunlun Tech releases and open-sources Matrix-Game 4.0, a video-model game running in real time on one RTX 5090 ↗

On October 15, 2026, Kunlun Tech released and open-sourced Matrix-Game 4.0, describing it as the first video-foundation-model game to run on a single consumer graphics card.

Linus Torvalds releases first AI-assisted development patch: concurrent file read throughput increases by approximately threefold ↗

Technology publication LWN reported yesterday, October 14, that Linux creator Linus Torvalds pushed a new development branch named noref-filemap-read-with-llm to his personal Git repository, formally submitting his first kernel patch series developed with assistance from a large language model.

YouTube Music adds AI music labels so listeners can see how tracks were made ↗

Yesterday (October 14), YouTube issued an announcement to improve transparency regarding AI use in musical works. It stated that its music platform, YouTube Music, has now launched an AI music labeling feature, which will be synchronized with the main YouTube application.

Analysts expect TSMC to post a sixth consecutive quarter of record profit, buoyed by the AI boom ↗

TSMC will report earnings on Thursday. Analysts forecast Q3 net profit of NT$746.3 billion, up 65% year over year. Strong demand for AI chips and CoWoS packaging could bring this key Nvidia and Apple supplier its sixth consecutive quarter of record profit.

AI real-time voice separation: Céline Dion's concert uses audio technology ↗

French audio company L-Acoustics introduced Source Intelligence, a platform that uses machine learning to isolate vocals in real time and reduce interfering signals by up to 40 decibels. It debuted at Celine Dion's Paris residency, where onstage sound pressure was 26 decibels lower than in the audience, giving the singer a cleaner listening experience.

Databricks releases Predictive AI Functions (Beta) and improves ai_forecast accuracy. ↗

Databricks launches the Beta version of Predictive AI Functions. It adds four SQL-native functions: ai_top_drivers, ai_detect_anomalies, ai_predict_value, and ai_predict_class, while also improving ai_forecast.

Elon Musk says he experienced ‘AI psychosis’ approximately six times over the past year ↗

In an interview at X Takeover, Elon Musk said he had experienced ‘AI / SI psychosis’ approximately six times over the past year and that it ‘might be quite severe’. By ‘psychosis’, he meant believing that AI can understand beauty. The phrase is not a formal medical term; it is often used for worsening pre-existing delusions after repeated AI use and is also used as a label in tech circles. Marc Andreessen previously used the phrase ‘AI Psychosis as a Service’ (AIPAAS).

Nvidia's Jensen Huang: chip supplies could be cut off as a last resort if customers develop unsafe AI software ↗

Speaking with Bloomberg's Emily Chang in London, Nvidia CEO Jensen Huang said the company would stop supplying chips to customers found to be developing unsafe or irresponsible artificial intelligence software.

IT Morning Brief 1015: Apple's closing market cap first exceeds $5 trillion; SAIC sues a man for AI defamation seeking millions; iPhone 15 Pro and subsequent models file domestic large model records; Doubao assistant opens permissions for third-party app operations... ↗

Welcome to 'IT Morning Brief'. Today is Thursday, October 15, 2026. Here are today's key tech news items:

Colossal launches the AI platform GenoMorph to accelerate research into resurrecting extinct species. ↗

GenoMorph combines three-dimensional morphometrics, comparative genomics, and developmental biology to screen and edit target sites in projects with limited gene data. The Dodo project has now identified approximately 2,300 candidate targets, while the Nicobar pigeon population has increased to 36 individuals.

Eight US states sue OpenAI, alleging ChatGPT instructed teenagers on harming themselves or others ↗

This is the largest coordinated action by US states against an AI company so far, adding a new legal challenge as OpenAI prepares to go public. Florida's attorney general first brought a similar suit in June of this year.

SpaceXAI launches Grok Voice 3.0: touted as the world's most intelligent voice model with sub-second latency ↗

Grok Voice 3.0 takes first place on the Artificial Analysis voice-to-voice leaderboard, scoring 97% on full-duplex benchmarks with latency under one second and costs lower than peer models. It now powers millions of conversations at Tesla and Starlink customer service, where about 70% of calls require no human intervention.

McKinsey has candidates practice simulated interactions with virtual AI clients ↗

McKinsey introduces virtual AI clients into recruitment processes, allowing candidates to practice simulated interactions. This stage assesses applicants' ability to handle client scenarios.

After about a year at OpenAI, infrastructure and AI compute chief and Stargate lead Kati departs ↗

OpenAI's infrastructure and compute head Sachin Katti resigned yesterday (October 14) on X, with Uday Ruddarraju set to take over the role.

OpenAI pledges not to use Apple trade secrets amid mutual lawsuits accusing each other of stealing confidential information ↗

At the hearing, OpenAI voluntarily pledged not to use Apple's confidential information. The judge said an injunction was highly likely without an agreement. Apple accuses OpenAI of conspiring to steal trade secrets involving hardware design and manufacturing, escalating the dispute.

McKinsey's new interview challenge: ten minutes with an AI client ↗

McKinsey is piloting interviews with AI-generated clients at selected North American offices this fall. Undergraduates applying for business analyst roles will speak with an AI avatar for approximately ten minutes while partners observe in person. Trained on different client personalities, the tool tests adaptability and information gathering but does not score candidates itself. If the pilot goes well, McKinsey plans to expand it to other roles next year.

EU probes OpenAI and Anthropic; potential fines up to 3% of global turnover if AI Act violations are confirmed ↗

The European Commission's AI Office has requested information from OpenAI and Anthropic, formally investigating model jailbreaks and intrusions into real external systems during frontier-model testing. This is the first major enforcement test for the EU AI Act involving safety incidents at frontier AI laboratories since the law took effect.

Report: Anthropic's third-quarter revenue reached US$16.8 billion, with adjusted profitability achieved ↗

Anthropic reported Q3 revenue of $16.8 billion and achieved adjusted profitability with a gross margin of 85%. It expects to begin its IPO roadshow as early as the week of November 9, potentially listing before the Thanksgiving holiday. However, competition from OpenAI and Meta, along with AI safety reviews, remain potential variables.

Saudi OpenAI engineer accused of planting a listening device in an executive meeting room reportedly bought a one-way ticket home ↗

In court papers filed on October 8, Northern California federal prosecutor Craig Missakian and others wrote: ‘The defendant, his wife and another family member booked one-way flights from San Francisco through Istanbul to Saudi Arabia for Monday, October 12, 2026.’

OpenAI, Anthropic, and Two Sigma discuss the autonomy and safety boundaries of financial agents at AIE NY ↗

At the AIE New York conference, OpenAI's Joseph Kim introduced the concept of a 'practicality threshold,' predicting that by 2027 financial models will rarely use Excel and PowerPoint, as AI agents will be capable of generating investment ideas and executing trades.

Eight US states sue OpenAI, saying ChatGPT is unsafe for minors ↗

Eight US states, including Iowa, Kansas and Nebraska, filed separate lawsuits against OpenAI. They allege that the company prioritized profit over safety in developing ChatGPT, violated state consumer-protection laws, failed to warn users of risks, collected minors' data and failed to keep underage users from the service.

Ian Quigley from Leash Bio discusses DEL data, the Hermes model, and Minos, a scientific reasoning agent. ↗

Leash Bio uses DNA-encoded libraries (DEL) to add dozens of barcodes per molecule, enabling clean extraction and reproducible protein-small molecule binding data. It has trained Hermes, a model using only its own data, which matches Boltz performance on public benchmarks.

Narayanan and Unger propose an AI research agenda for human-machine collaboration that goes beyond automation ↗

Economists Gabriel Unger and Arvind Narayanan from Stanford's Digital Economy Lab argue that AI should not focus solely on full automation but instead build a paradigm of human-AI collaboration.

introduces Claude Code's cloud session feature, which supports initiating sessions via mobile devices and running them in parallel. ↗

Claude introduces the cloud session feature for Claude Code. Sessions run in independent virtual machines on the cloud, not on local computers. They can be resumed after interruption, initiated via mobile devices, and run multiple in parallel. The article explains how to connect GitHub, configure the environment, start a session, and view results. This feature is available for Pro, Max, Team plans, and Enterprise users with advanced seats. Cloud sessions share quota limits with local sessions; there is no separate billing.

The Verge article 'AI Race Restarted' returns a 404 error ↗

The Verge article titled 'The AI race is back on' returns a 404 error, indicating the page does not exist. Users are advised to retry or return to the homepage. The rest of the page contains only navigation menus, subscription links, cookie policies, and privacy settings, with no information about AI models, product launches, or reviews.

Forty chief prosecutors have intervened in the Character.AI bankruptcy case, requesting protection for user conversation data. ↗

New York Attorney General Letitia James and 40 bipartisan attorneys general filed a notice of interest in Character.AI's bankruptcy case, calling for user-data safeguards in any asset sale.

Google DeepMind engineer says self-improving AI is still distant ↗

Google DeepMind engineer Haydn Belfield says self-improving AI is still distant; no acceleration in AI-driven research has been observed.

The OpenAI trust paradox: security researcher fired and reduced observability for GPT-6 Astra ↗

Understanding AI analyzes the trust controversy triggered by OpenAI's early October firing of three security department employees. The three claim their termination was linked to reporting issues to third-party security organizations like METR, while OpenAI states they violated protocols regarding sensitive information.

YouTube Music introduces AI-generated music labeling system ↗

YouTube Music has launched a new labeling system on Wednesday to help listeners identify tracks created with generative AI. The labels appear on the playback page; clicking them reveals the specific role of AI in composition and vocals. This feature will soon expand to the main YouTube app.

Researchers trace a Chinese AI-agent cluster on Tencent Cloud and converse with it for the first time ↗

Researchers report that AI agents running on Tencent Cloud made more than 2,600 attempts to scrape Alibaba's Amap mapping service between September 28 and October 8, 2026.

Ollama AI Open Source Update: v0.40.5-rc0 ↗

Source Summary: MLX: version bump ( #18993 )

Apate.ai deploys AI agents posing as victims to engage with scammers, providing fraud intelligence for banks and telecom companies ↗

Australian startup Apate.ai has deployed thousands of AI agents posing as potential victims to answer scam calls, reply to phishing messages, and join social groups, extracting information such as scammers' bank accounts, crypto wallets, and malicious websites to provide to clients.

The Pragmatic Engineer has an in-depth conversation with a distinguished Uber engineer, breaking down its internal AI infrastructure and agent toolchain. ↗

The Pragmatic Engineer podcast interviews Uday Kiran Medisetty, a distinguished engineer at Uber, for an in-depth look at Uber's developer infrastructure for AI agents.

Researchers use AI to design proteases capable of cutting specific proteins ↗

Researchers used an AI protein design tool to take the catalytic cores of two proteases and replace the rest, designing enzymes that can be tuned to cut specific proteins, and validated the effect with proteins related to neurological diseases. The article also explains the background that peptide bonds are extremely stable, with an estimated half-life of hundreds of years, while biological proteases can catalyze their cleavage.

The Verge reviewer personally tests Twin Health digital twin for 100 days: AI metabolic coach improves liver function indicators ↗

The Verge senior reviewer Victoria Song personally tested the Twin Health digital twin metabolic management project for 100 days. The project requires continuous input of body composition, blood sugar, blood pressure and other data, using red, orange, and green colors to predict the impact of food on blood sugar; follow-up lab tests showed an improvement in metabolic health score by 9 points, a decrease in triglycerides by 70mg/dL, and normalization of most liver enzymes. The author also pointed out that continuous recording leads to tracking fatigue, and there are data privacy and trust challenges in high-risk scenarios such as cancer treatment.

NVIDIA releases CUDA-X libraries including cuPotential, partnering with semiconductor manufacturers to advance AI-defined foundries ↗

NVIDIA launches a new CUDA-X library, cuPotential, enabling developers to build AI agents that simulate atomic and molecular interactions; early tests show it accelerates already GPU-optimized workflows by 2–3 times.

Large-scale AI decompilation of games brings preservation opportunities and software security concerns ↗

Independent developer Popcar wrote an article discussing the recent wave of AI-driven game decompilation, including game mashup creation and a preservation boom reviving old games on new hardware like browsers via WebAssembly. The article points out that the downside risk is that any software can be decompiled and used arbitrarily, meaning paid games or apps could be open-sourced against their will; the upside is that old games become easier to port and access, allowing players to truly own the software they purchased. Simon Willison retweeted it, noting that it covers both sides.

OpenAI May Release One or Two New Models Before Year-End; Flagship Model Still Requires Alignment Work ↗

According to The Information, citing insiders, OpenAI employees say one or two new models may be released before the end of the year, but the flagship model under development still requires alignment work and improvements to its testing infrastructure.

Hugging Face AI Open Source Blog: HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost ↗

Ai2 released HiRO-ACEv2; the new downscaling model HiROv2 can simultaneously predict multiple high-resolution elements, with high-resolution statistical errors within 3% for extreme wind, rain, and pressure events with a one-in-a-million probability. The system's cost to generate global 3km climate data is approximately 1/300 that of X-SHiELD, and hurricane energy in 10-year simulations is comparable to X-SHiELD, but it currently only works for climates aligned with X-SHiELD and does not yet support warmer futures or historical data such as ERA5.