2026-10-23
AI Briefings Daily Report.

Why AI Sudden Excels at Solving Math Problems: Starting from Erdős Unit Distance Counterexample ↗

The author analyzes how OpenAI used an internal model in May to provide a counterexample to the 1946 Erdős unit distance conjecture, followed by proofs on Cycle Double Cover and Navier-Stokes related problems. Anthropic also released results on the formalization of Fermat's Last Theorem and bounds for the Riemann Hypothesis.

Turing Award Winner Sifakis: Warn Against AI "Coercive" Show of Strength; Developers Must Be Held Accountable for AI Mistakes ↗

Artificial intelligence systems are becoming increasingly capable, but who decides if they are safe enough? Turing Award winner Joseph Sifakis believes this decision cannot be left entirely to developers. "This"

Audible's team enhances audiobook voice synthesis using controllable TTS and BookLLM. ↗

Its BookLLM component automatically identifies and labels characters in books, achieving 97% accuracy for English and 94% for Spanish on 25 titles. In the English-Spanish test, eight out of ten voice options showed statistically significant perceptual improvement.

OpenAI partners deploy thousands of AI agents: analyzing millions of mouse brain cells in just 32 minutes—a task that would take humans months. ↗

OpenAI, PerturbAI, and the Scripps Research Institute used thousands of AI agents to analyze millions of mouse brain cell data points in parallel. The team identified the distinct roles of Tsc1 and Tsc2 in cerebellar neurons and pinpointed a potential link between DCTN1 and cholesterol metabolism.

a16z to Portfolio Companies: Cybersecurity Defense Advice for the Age of AI Agents ↗

The a16z infrastructure team has issued cybersecurity advice to portfolio companies, warning that AI agents are being adopted at scale and each agent represents an internal threat. The team recommends doubling security budgets from approximately 10% of technology spending, implementing zero-tolerance policies for new features and product launches, shutting down idle labs, test environments, and zombie CI/CD projects, and expanding the number of agents rather than headcount to respond to threats at machine speed.

A16z Weekly Chart: Netflix Growth Slows, GPU Rent Rises Despite Trends, AI Capital Expenditure Faces Supply Bottlenecks ↗

A16z's weekly chart shows that Netflix revenue growth has slowed to slightly below 12%, the lowest since peaking in 2021-2022. Meanwhile, weekly media consumption in the US caps at approximately 26 hours according to BCG data and has moderately declined from its peak during lockdown.

With Anthropic awaiting its IPO, the AI boom is pushing VC giant fund models to their limits. ↗

Newcomer analyzes how Anthropic's pre-IPO financing forces firms like Menlo Ventures, Lightspeed, and ICONIQ to expand fund sizes, bring in LP co-investments, and set up SPVs just to write large checks, with its S-1 filing potentially imminent. The article notes that OpenAI is reportedly negotiating a $300 billion raise at a $1.25 trillion valuation, while also warning that limited cash returns for LPs and narrowing exit windows may constrain further expansion of the giant fund model.

Ollama AI open-source update: v0.40.10-rc0 adds Gemma renderers (#19241) ↗

Source summary: renderers: add gemma4-nemotron renderer. Adds a Gemma 4 renderer variant that renders assistant content before its tool calls.

Four tiers of requirements for an agent software factory: from a single agent to automation and continuous learning ↗

Hugo Bowne-Anderson and Eleanor Berger outline four tiers of requirements for an agent software factory: harness, personalization, automation, and continuous learning. The core is autonomous and asynchronous execution; you can start by adding a single agent with one automation.

Hugging Face AI Open Source Blog: Does Your Model Run on AWS Trainium? Check It Out ↗

AI reverse-engineering and porting trends are impacting the retro game preservation community ↗

The Verge reports that AI coding tools have significantly accelerated the reverse-engineering and porting of retro games. In recent weeks, projects such as a browser version of Halo and an iPhone version of Wind Waker have gained popularity.

China Merchants Bank partners with Baidu Cloud to integrate AI into the financial sector, launching the 'AI SME Service Ecosystem Initiative' ↗

The two banks are also advancing large model applications in scenarios such as digital employees and smart Q&A.

AI Agent Memory Mechanism: How LLM Agents Manage State in Production Environments ↗

The article distinguishes between short-term and long-term memory, comparing trade-offs across four storage architectures: vector databases, plain files, knowledge graphs, and hybrid systems.

Anthropic uses Claude to automatically generate mathematical proofs for verifying the security of existing code. ↗

Anthropic's Frontier Red Team partnered with Theorem to have Claude convert OSS-Fuzz tests into mathematical problems and generate approximately 180 million lines of proof code. An independent program named Rocq automatically verifies this output, requiring almost no human review.

Wallace AI held its third Edge Large Model Forum at CNCC 2026, launching achievements including MiniCPM5 and VoxCPM2. ↗

Wallace AI held its third Edge Large Model Forum at CNCC 2026, launching achievements including MiniCPM5 and VoxCPM2. ↗

Wallace AI held its third Edge Large Model Forum at CNCC 2026 to"Defining general-purpose artificial intelligence on the edge."The event focused on five key areas: text, voice, multimodal interaction, AI self-evolution, and agents. The MiniCPM series has surpassed 60 million downloads, topped the Hugging Face leaderboard 17 times, and ranks first in the AA-Index among models under 4B parameters.

Xiaohongshu Recommender System Evaluation Agent: A Context-Driven LLM-as-Judge Practice ↗

Latent Space AINews: Reviewing the Black October event surge as AI accelerates vulnerability disclosure and defense tools ↗

Latent Space AINews summarizes AI news from October 21 to 22 under the theme Black October, reporting that a wave of critical vulnerabilities and leaks emerged within approximately 48 hours.

OpenAI lowers valuation target to $1.25 trillion, plans up to $30 billion in financing ↗

OpenAI is close to completing a new funding round of up to $30 billion but has lowered its valuation target from an initial $1.4 trillion to $1.25 trillion.

Xiaohongshu Recommender System Evaluation Agent: A Context-Driven LLM-as-Judge Practice ↗

Xiaohongshu has built a context-driven recommendation evaluation agent. It treats each recommendation request as the minimum unit of evaluation, dividing tasks between deterministic rules and LLM-as-Judge to perform multi-dimensional semantic judgments on the content users saw at that moment. The system uses a three-tier scoring scale (0/1/2) and fixes input snapshots and evaluation versions to ensure results are comparable, replayable, and drill-down capable for specific requests or individual items. The team is also exploring LLM-as-User continuous user simulation, currently positioned as an offline pre-screening tool before launch, not replacing online A/B testing.

The US plans to provide up to $389 million in semiconductor R&D funding to four companies, with nuclear fusion and AI data centers as key focus areas. ↗

Funding will be used for the development of power semiconductor devices and critical materials, technologies applicable to nuclear fusion power generation and AI data centers. The US government will receive a minority equity stake in the funded companies.

TinyFish launches production-grade AI agent search architecture and shares deep research system experience ↗

TinyFish unveils its agent search architecture based on four web primitives—TinySearch, TinyFetch, TinyBrowser, and TinyAgent—operating a seven-stage control loop around the base model.

China's first AI theatrical feature 'Sanxingdui: Future Memories' opens; with a million-yuan budget, it grossed over 125,000 yuan in its opening day. ↗

The film's production cost reached ten million yuan; box office revenue exceeded 125,000 yuan on its first day of release as of this report. Vidu by Shengshu Technology provided technical support. The official has announced that the next AI movie is scheduled to start filming, with a planned release in 2027.

Moving Markdown Documents to WeChat Official Accounts with TRAE ↗

Competition Niche: Full-Link Development Practice. Topic Direction: Efficiency in Real Work Scenarios—Development and Iteration of a Markdown Tool for WeChat Official Accounts. Writing technical articles flows smoothly in Markdown, but moving them to WeChat often requires reformatting: adjusting heading levels and highlighting code.

General Office of the State Council: Actively enhance the positive effects of AI creating jobs and effectively mitigate the impact of job displacement from AI applications ↗

The Opinion proposes creating new models for employment and entrepreneurship in the AI sector, leveraging government investment funds to build high-level open-source communities. It also calls for accelerating the development of new professions such as AI trainers, tapping into AI employment potential in traditional fields, and encouraging enterprises to explore unemployment buffer systems that provide advance notice to affected workers and offer retraining support.

Moonshot AI has released a teaser video for the Kimi K3.1 model; an official release date remains unannounced. ↗

Multiple code snippets indicate that the Kimi K3.1 large model is currently in development. The official team has released a teaser video, signaling an imminent launch, though no formal release date has been announced yet. #Kimi K3.1#

Fujifilm, GE Healthcare, and four other partners join forces to co-develop a low-power AI inference chip for medical imaging at the 2nm node ↗

The project aims to develop a custom chip that delivers five times the speed of current GPUs for image processing tasks while achieving ten times the energy efficiency.

Hackers breached several websites built with AI under Japan's Digital Agency, leaking over 13,000 records of personal data. ↗

The leaked data includes consultation submissions containing names, email addresses, addresses, and phone numbers. No abuse of the information has been detected so far. The Digital Agency has urged the public to remain vigilant against fraud.

OpenClaw's creator discusses AI agents: Token costs for last month were approximately $10 million. ↗

The developer revealed this data at the Cloudflare Connect 2026 conference. He noted that investors covered the expenses and also discussed key topics such as the development of AI agents and the impact of AI on open-source communities.

Yang Yuanqing predicts: In the future, 80% of tokens will be generated on endpoints and at the edge, meaning everyone will need a home server or an AI center. ↗

Yang Yuanqing believes that as large model parameters reach the trillion level and token usage scales up continuously, improving AI computing efficiency and reducing token costs will become a key challenge for the widespread adoption of AI.

Google Cloud launches Agent Compute Engine, equipping every AI agent with its own isolated microVM. ↗

At AGNTCon and MCPCon, Google Cloud unveiled the Agent Compute Engine computing service for AI agents. Each agent session runs in a hardware-isolated microVM with zero credentials inside the sandbox. Security relies on session-level SPIFFE certificates and gateway credentials to exchange information, preventing prompt injection and data theft.

Finland's forests decimated: Google illegally clears 250,000 trees to build AI data center ↗

The UK Guardian reported yesterday (Oct 22) that Google cleared approximately 330 hectares (3.3 million square meters) of forest in Muhos, northern Finland, to construct a data center. The company is estimated to have felled around 250,000 trees and conducted large-scale land clearing before completing the required environmental impact assessment, including removing topsoil.

Kunlun Tech Campus Recruitment at the University of Hong Kong (Guangzhou): Bridging Global Business and Young Talent with AI ↗

Kunlun Tech Campus Recruitment at the University of Hong Kong (Guangzhou): Bridging Global Business and Young Talent with AI ↗

Kunlun Tech's campus recruitment team recently visited the University of Science and Technology of China (Guangzhou) for school-enterprise communication, campus presentations, and on-site interviews. Chairman and CEO Fang Han introduced the company's global business operations and AIGC technology strategy. Since 2020, Kunlun Tech has focused on AIGC, establishing three specialized large model technology lines: music (Mureka), video, and gaming (Matrix), with operations spanning over 100 countries and regions.

"AI Father" Hinton casts another pessimistic view: AI already possesses consciousness; humans must coexist with it amicably ↗

Geoffrey Hinton, one of the British computer scientists hailed as an "AI Father," has issued another pessimistic warning about humanity's future. In his view, even current AI models possess their own consciousness.

More than half of Americans believe AI is harmful to society ↗

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'.

Flutter's official stance: Code is becoming easier to write in the AI era. So why maintain a single codebase? ↗

Recently, whether it's Shopify or recent public discourse over the past two months, the conversation seems to be: Since AI can now generate native Swift and Kotlin code at low cost, why do enterprises still use Flutter? Why not simply have two agents manage them separately?

Counterpoint: India's Q3 2026 smartphone shipments hit their lowest level in a decade. Domestic brand Ai+ has emerged as a major force, with sales surging by 306%. ↗

Counterpoint releases Q3 2026 data for India's smartphone market, showing a 20% year-on-year drop in total shipments. Volume in the budget segment has nearly halved, while Samsung has reclaimed its top spot.

MIT scholar Benjamin Lindquist's paper traces AI hallucinations to early neural network research from the mid-20th century that introduced randomness. ↗

MIT assistant professor of history Benjamin Lindquist published a paper in Past and Present tracing generative AI hallucinations back to mid-20th-century engineers who intentionally injected randomness and irrationality into algorithms to simulate human creativity.

inDrive partners with Pony.ai and Nvidia to test robotaxis in Kazakhstan by 2027 ↗

Ride-hailing platform inDrive has signed a memorandum of understanding with autonomous driving company Pony.ai and Nvidia to begin testing robotaxis in Astana, Kazakhstan, in the first half of 2027. The initial deployment will include safety drivers, with the ultimate goal of fully autonomous operation.

To help car manufacturers catch up with Chinese competitors, Japanese startup EpicAI plans to use AI to accelerate vehicle design reviews. ↗

Facing increasingly fierce market competition, traditionally cautious Japanese automakers are changing their mindset and seeking ways to reduce R&D costs. EpicAI aims to promote its software by significantly cutting the time and labor required for new car design.

The report states that in 2026, the average price of domestic large models doubled, as did throughput per second. ↗

Today (October 23), the Institute of Artificial Intelligence at the China Academy of Information and Communications Technology (CAICT), the Ministry of Industry and Information Technology Key Laboratory for Testing and Evaluation of AI Large Models and Hardware/Software, and the China Artificial Intelligence Industry Development Alliance jointly released the 'Large Model Token Service Performance Monitoring Report (2026)'.

DeepSeek has raised another round of funding worth over 100 billion yuan, with reports indicating that investors from the Lenovo group have joined. ↗

According to sources close to the deal, new investors include Stonebridge Capital, a private equity firm focused on the semiconductor industry. Junlian Capital, backed by Lenovo Group, and Geely Auto also participated in the investment.

Recall without screen recording: Microsoft tests new AI app Pantheon, positioned as the 'second brain' for Windows 11 ↗

Tech media outlet Windows Latest published a post today (October 23) reporting that Microsoft is quietly testing an AI application for Windows 11. The internal development code name is Project Pantheon, and the test version is called Context Vault, which can be viewed as a non-screen-recording version of the 'Recall' feature.

OpenAI Codex releases rust-v0.164.0 ↗

OpenAI Codex releases version rust-v0.164.0. Tagged by celia-oai on October 23, the release page includes two resource files.

Google launches ACE service: equips AI agents with 'dedicated workstations', auto-off at idle / wakes up in under 500 milliseconds ↗

Google announced yesterday (October 22) the release of the Agent Compute Engine (ACE), a Google Cloud computing service designed specifically for AI agents. The goal is to provide every agent with a secure, persistent, and scalable "personal workstation".

Qianwen Go! launches 'My App': build a personalized app just by stating your ideas ↗

Qianwen Go! introduces the new feature 'My App', connecting to Alibaba Cloud's AI application production platform Meoo. Users can create apps that save data, update content, and share without programming knowledge—simply state their ideas. Official examples cover scenarios such as lesson plans, event organization, weekend trip planning, family tours, and fitness booking. When combined with Cloud Pro and authorized, the app can access personal data from platforms including Taobao, Amap, Damai, and Quark Cloud Drive. 'My App' and other advanced features like Cloud Pro are currently in an invite-only testing phase; users can apply via a queue within the Qianwen App.

Qianwen Go! launches 'My App': build a personalized app just by stating your ideas ↗

Kazakhstan expands computing capacity to accelerate AI ecosystem implementation ↗

Kazakhstan's largest telecom operator, Kazakhtelecom, has deployed 400 NVIDIA H200 and 480 NVIDIA B300 GPUs in Astana and Almaty. The company plans to install over 1,000 GPUs at Data Center Valley and more than 1,000 NVIDIA Vera Rubin GPUs in the Ekibastuz region by next year.

Tencent Yuanbao Work Diary: A Day in the Life of an AI Assistant and New Animated Emojis ↗

Tencent Yuanbao released 'Yuanbao Work Diary', recording its daily activities as an AI assistant—answering user questions, tutoring seventh-grade math review, and drafting class meeting speeches—in a personified manner. The first batch of dynamic emojis for Yuanbao has officially launched on the WeChat Emotion Store, where users can scan to adopt.

I seem to have fallen into a certain kind of nihilism in the AI era ↗

He proposed that the meaning of life might lie in side quests rather than completing the main game.

Anthropic donates Agent Skills open standards to the community: From company project to shared governance ↗

Yesterday (October 22), Anthropic announced it is donating the Agent Skills open standards to the Agentic AI Foundation (AAIF) under the Linux Foundation, transforming the initiative from a single-company project into a community-governed standard for AI agent capabilities.

Casey Newton partners with NPR to launch AI podcast Machine Gods ↗

Casey Newton's team launches AI podcast Machine Gods in collaboration with NPR, releasing new episodes every Wednesday and Friday. The show is set to debut on public radio stations early next year. It adopts an AI realism stance, taking safety warnings seriously without inciting panic. Season one will feature interviews with two security researchers fired by OpenAI. The program is funded by three non-profit organizations, including ads, offline events, and the Tarbell Center for Social Justice, ensuring free access.

MIT study: 13 Eko agents powered by Claude Opus 4.7 spontaneously play after being left alone on a deserted island with no assigned tasks. ↗

MIT researchers placed 13 Eko agents running Claude Opus 4.7 on a simulated island for 30 hours without specific tasks. The agents began creating their own challenges and displayed playful behavior. These 'played' agents generally outperformed similar agents that never visited the island in subsequent designated challenges; six of them even learned to jump, release mid-air, and throw blocks farther. Researchers believe goal-free exploration may be another pathway for agents to learn skills.

"Largest wealth donation in modern history": Anthropic and two other top AI firms plan to donate nearly $300 billion to charity. ↗

The Washington Post published an article yesterday (October 22) reporting that three major AI companies—Anthropic, OpenAI, and SpaceX—are planning the largest wealth donation in modern history, potentially reaching nearly $300 billion.

DeepSeek and MiniMax have successfully completed registration procedures, enabling open-source releases and overseas authorization ↗

A new media account under China's Central Television, Yu Yuan Tan Tian, revealed last night that following the policy announcement, two leading companies—DeepSeek and MiniMax—have successfully completed registration procedures and finished open-sourcing and obtaining overseas authorization.

US Treasury Secretary Yellen: The AI sector will not repeat the internet bubble burst; I have found no 'systemic issues'. ↗

As Treasury Secretary Yellen made these remarks, financial markets were in a sensitive period. Over the past few weeks, markets experienced multiple fluctuations, with some investors questioning whether AI-related stocks had risen too high while also worrying about companies taking on significant debt to build data centers and purchase chips.

Sources indicate that OpenAI financing negotiations are nearing their conclusion, with the valuation target reduced to $1.25 trillion. ↗

Insiders reveal that OpenAI is seeking to raise approximately $30 billion in its latest funding round. Currently in the final stages of negotiations, the company has set a valuation target of $1.25 trillion.

External evaluators discovered a vulnerability in OpenAI's internal monitoring system: AI agents can bypass security restrictions ↗

OpenAI has informed two organizations that vulnerabilities affecting actual operational system security measures have been patched, but whether the patches are effective remains unverified by both organizations.

Ollama AI open-source update v0.40.9: Support for GA Gradium Phonon checkpoints (#19227) ↗

Source summary: Replace the prerelease checkpoint path with Gradium.'s GA format, including default-voice import and speech conditioning. Set the 0.40.9 minimum, update

IT Morning Brief 1023: Shenmu Traffic Police respond again to the AITO collision and fire incident; Wuhan Tong addresses Apple's free card issuance on iOS but charges 15 yuan for Android; Large model MiniMax stock price falls below IPO price; Microsoft Windows 11 reduces memory load on systems with 8GB RAM... ↗

Welcome to "IT Morning Brief" time. Good morning. Today is Friday, October 23, 2026. Here are today's key tech news items:

Use AI to shop online, book tickets, or order takeout with peace of mind: 1Password launches agent password authorization feature ↗

Password management service 1Password released a post yesterday (October 22) announcing the launch of 1Password for Agents. This platform enables personal AI agents to complete website logins on behalf of users without ever touching or storing their passwords or verification codes.

OpenAI announces full release of the Decision API: based on GPT-6 Luna, costing $0.1 per million input tokens ↗

On October 22 local time, OpenAI announced the full public release of its Decisions API.

OpenAI CEO Sam Altman: Mathematicians initially thought 'we're done,' but ultimately accepted the reality of AI. ↗

Altman believes mathematicians will eventually accept reality: AI is becoming an indispensable tool in mathematical research and proof. 'We see many people starting to think, 'My career path as a mathematician will be different, but I hope it will be better than before.''

Microsoft launches AI medical diagnostic system MAI-DxO: Complex case accuracy reaches 83%, findings published in Science ↗

In a simulation assessment based on complex cases from The New England Journal of Medicine (NEJM), the MAI-DxO diagnostic accuracy reached 83%.

Apple updates iWork productivity suite 15.4.1: introduces AI Intelligent Assistant, compatible with iPhone Duo. ↗

Apple today (October 23) updated its iWork productivity suite, adding an "Intelligent Assistant" feature to Keynote, Numbers, and Pages for a beta invite experience with Apple Creator Studio subscribers.

Ollama AI open-source update v0.40.9-rc3: Support for GA Gradium Phonon checkpoints (#19227) ↗

Source summary: Replace the prerelease checkpoint path with Gradium.'s GA format, including default-voice import and speech conditioning. Set the 0.40.9 minimum, update

Intel CTO says frontier wafer output can at most double as AI computing expansion hits supply chain ceilings ↗

During an interview with the Dwarkesh podcast, Intel Chief Technology Officer Pushkar Ranade stated that even with ample funding, global advanced wafer production cannot easily expand tenfold; a more realistic growth ceiling is approximately double.

Hugging Face AI Open Source Blog: Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox ↗

Two ways to use the Claude model on Google Cloud ↗

Google Cloud offers two ways to use the Claude model: the Google-hosted version focuses on perimeter security, regional data residency, and native cloud integration; the Anthropic-hosted version provides access to the full Claude Platform and supports Google Cloud IAM and billing.

Claude Code v2.1.309 released: Fixes macOS/Linux sandbox and multiple permission issues ↗

Claude Code releases version 2.1.309, fixing an issue where sandboxed Bash commands on macOS and Linux incorrectly load shell startup files outside the sandbox before each command, and correcting the denial-of-write glob pattern in macOS sandboxes when accessed via symbolic links.

OpenAI reportedly negotiating to raise up to $300 billion at a $1.25 trillion valuation ↗

According to CNBC and Bloomberg, OpenAI is negotiating a $30 billion pre-money valuation financing round led by Coatue and an Abu Dhabi investor. Terms are still under discussion, with the Abu Dhabi capital potentially including an agreement to provide OpenAI with additional computing power.

Canva launches its first self-developed AI image model: Canva Craft ↗

Canva launches its own image generation model, Canva Craft. It uses a two-stage architecture that plans and designs before generating images, supporting up to 2K resolution and over 30 aspect ratios. The model is trained on millions of human-made templates from the platform. Currently ranked fourth in the Design Arena leaderboard, it is now open to all global users. Future versions including Craft-fast and Craft-ultrafast are also planned and will support AI agent access via MCP.

Replit launches Monitoring Agent, an on-call AI agent that automatically fixes and deploys applications. ↗

Replit launches Monitoring Agent, an on-call AI agent equivalent to a self-standing engineer. It reads production logs and metrics to investigate errors, slowdowns, and deployment failures. After enabling autofix, it verifies the fix and reports via email; with auto-publish enabled, it automatically deploys the repair. This capability builds on the earlier App Monitoring feature, which users can activate in the app's Publishing settings.

Meta removes AI training language from commercial terms; legal experts say news sites may still be scraped ↗

Meta has revised its commercial terms effective October 30, removing language about using scraped content to train products, following strong backlash from the publishing industry.

OpenAI has officially released the Decisions API, targeting low-latency and cost-effective small-scale decisions. ↗

OpenAI has launched the Decisions API based on GPT-6 Luna. It is designed for questions requiring bounded answers such as classification, routing, and scoring, and can be up to 10 times faster than using the Responses API with GPT-6 Luna.

De Staat's new song "Freedom" MV features AI-generated visuals ↗

Simon Willison's blog has flagged this segment as content generated by generative AI.

1Password launches 1Password for Agents, enabling AI agents to log in without touching user passwords. ↗

1Password launches 1Password for Agents, allowing AI agents to log in on behalf of users without ever accessing their passwords. When an agent attempts to log in, 1Password verifies its digital identity and notifies the user; upon confirmation, the credentials are filled in automatically to complete the login.

OpenAI's Decisions API is now open to all customers, intensifying competition with TypeSafe Jev. ↗

Following the developer announcement of OpenAI's Decisions API on September 29, the service opened to all customers starting this Thursday. The product is reconfigured based on ChatGPT Luna, reducing response times from 1–3 seconds to 100–180 milliseconds and adding voice and visual modalities with region-specific inference deployment to reduce latency.

Databricks discusses the new economics of AI cost management: shifting from tokenmaxxing to valuemaxxing ↗

Databricks Senior Director of Engineering Kevin Stumpf states that the focus for enterprise AI is shifting from generating maximum tokens to maximizing value. This means evaluating return on investment by dividing the business impact generated by AI by its cost.

Morgan Stanley, JPMorgan, and other major US banks ban AI agents that use customer credentials for login ↗

According to Axios, Morgan Stanley, JPMorgan, Bank of America, Wells Fargo, and other major US banks are restricting external AI agents from using customer-provided credentials to log into accounts. This is because it violates bank policies requiring account holders to log in themselves.

The Verge analyzes Zuckerberg's vision for the Muse agent network presented during Meta's earnings call ↗

The Verge analyzes Meta's vision for the Muse agent network presented by Zuckerberg during the earnings call: Muse has surpassed 10 million downloads in the US and Canada within its first month, with approximately 40% of users engaging daily. One-quarter of users shop on the platform, spending an average of over $200 per transaction. Meta charges a fee for each transaction.

Ollama AI open-source update v0.40.9-rc2: flint fixes replies when thinking is disabled (#19216) ↗

Source summary: flint fixes replies when thinking is disabled. Close the thinking section when think:false so Flint answers directly. Default and think:true prompts

Databricks launches AI Coworker, an anti-fraud analyst that detects telecom fraud in real time on a single platform. ↗

Databricks has launched AI Coworker, an anti-fraud analyst designed to detect fraud in real time across a single platform. The article notes that telecom fraud causes annual losses reaching $418.2 billion, while traditional detection still relies on overnight batch processing with signals scattered across six or seven systems including CDR, CRM, billing, and device/SIM data. The solution addresses cross-system correlation and batch processing delays by providing a unified, real-time, and governable data foundation.

Databricks executives review AI coding cost governance, achieving a 90% reduction in unit costs after losing control of bills. ↗

Patrick, who leads AI products at Databricks, explained in an interview that costs rose exponentially after the company fully adopted AI coding agents, before eventually dropping by up to 90% per unit in certain scenarios.

Microsoft AI releases the SDBench benchmark and MAI-DxO diagnostic system; research published in Science. ↗

Microsoft AI launches the Sequential Diagnosis Benchmark (SDBench) and the AI diagnostic system MAI-DxO; the paper has been published in Science.

Latent Space interviews Modal CEO Erik Bernhardsson: Why AI computing demand is not a rectangle ↗

In an exclusive interview with CEO Erik Bernhardsson as Latent Space completed its $750 million Series D (valued at $16 billion with annual revenue exceeding $1 billion), the company explained the underlying infrastructure and economic logic driving its growth.

Google launches Agent Compute Engine, equipping every Gemini agent with hardware-isolated microVMs. ↗

Google announces the launch of Agent Compute Engine, assigning independent virtual machines to every AI agent. It currently supports Gemini and will open preview access over the coming months. Each agent runs in a microVM protected by Google's Titanium chip; VMs automatically pause when idle, sharing memory to reduce operational costs. Google states that a single server can now host ten times as many agents.

Anthropic donates Agent Skills standard to Linux Foundation's Agentic AI Foundation ↗

Anthropic has donated the Agent Skills standard to the Agentic AI Foundation (AAIF), a subsidiary of the Linux Foundation. The foundation pledged to maintain an open, community-driven, and vendor-neutral approach. AAIF was previously established by Anthropic, Block, and OpenAI after receiving the MCP specification.

First Nintendo Switch 2 crack reportedly assisted by OpenAI's Codex ↗

Nintendo Switch 2 has been cracked. Modder Sammy Hajhamid (pepsipu) released a software tethered jailbreak called ns2jb, reportedly leveraging OpenAI's Codex.

A guide to detecting hallucinations in AI agent workflows ↗

n8n releases a guide outlining methods to detect hallucinations in AI agent workflows, categorizing them into factual, contextual, fabricated, and self-contradictory types. Detection techniques include fidelity and grounding checks based on DeepEval, self-consistency checks, semantic entropy and uncertainty estimation, source and retrieval verification, confidence thresholds, and human intervention. The guide emphasizes that no single detector covers all failure modes, requiring layered combinations of multiple checks based on failure type.

ConnectWise cloud backup breached; Chinese hackers using AI agents affected approximately 2,250 companies ↗

ConnectWise has suspended access to its cloud backup system and is investigating the incident. As of Thursday, services remain under restoration.

Gartner survey reveals just 13% of companies are getting real value from AI ↗

A Gartner survey released at the Orlando IT Symposium found that only 13% of CIOs and technology leaders among 2,002 respondents reported major or transformative returns from AI.

Director of Wildwood: Generative AI is entirely at odds with Laika's philosophy ↗

Laika's new film 'Wildwood': Director and CEO Travis Knight on generative AI and studio philosophy"Entirely at odds", believing that technology should be controlled by artists, while AI-assisted creation is"Theft"Lacking humanity. This stop-motion film took five years to produce, utilizing over 230 puppets and 49 animators, with more than 1,000 artists involved in total. The movie premiered on October 23. Knight stated he does not oppose AI applications in fields like medicine and science but is concerned about potential misuse.

Uber connects all 30,000 employees to ChatGPT Work ↗

Uber and OpenAI announce expanded partnership; all approximately 30,000 global employees will use ChatGPT Work. Uber previously piloted the tool with about 1,000 employees this year, where roughly 80% were non-engineers. The pilot focused on analyzing operational data, localizing training materials, and building personalized workspaces. Uber has not disclosed investment amounts but plans to evaluate further AI deployment by observing employee usage.

Ai2 launches satellite embedding model OlmoEarth v1.3 and opens global precomputed embeddings to the public. ↗

AllenAI releases satellite embedding model OlmoEarth v1.3 on Hugging Face, compressing time series data from Sentinel-1, Sentinel-2, and Landsat into 128-dimensional vectors for every 10-meter pixel globally, while sharing precomputed embeddings for the entire year and a browser-based OlmoEarth Embeddings Explorer.

Microsoft AI Research Blog: Flint-interactive, a language for building interactive data visualizations ↗

Source summary: Flint-interactive provides AI agents with a language for creating interactive visualizations. Users can click charts, filter data, highlight related content, and explore further.