2026-10-17
AI Briefings Daily Report.

Meta removes language from its terms that could be interpreted as permitting news websites to scrape content for training AI models. ↗

Following controversy, Meta has updated its commercial terms to remove language that experts interpreted as permitting the scraping of news content for AI training and overriding robots.txt restrictions. Cornell law professor James Grimmelmann and others previously interpreted these terms in that way. In response, Meta spokesperson Daniel Roberts stated the provisions apply only to advertising and commercial products, and Meta's use of web crawlers remains unchanged. The update will not grant Meta new rights to override content creators' preferences.

Claude from Anthropic cracks the mystery of the NSA's 25-year-old unsolved ECDSA seed puzzle in under 30 minutes ↗

Claude from Anthropic cracked a cryptography puzzle that had been unsolved for 25 years in under 30 minutes. The ECDSA elliptic curve seed standardized by NIST in 2000 was provided by the NSA, with an unknown origin, leading to long-standing suspicions of backdoors.

"Wish-list" programming capabilities upgraded! DeepSeek launches: Open-source proprietary search API with calls costing only 5 yuan per thousand requests. ↗

"Big things are coming." A sudden outage a few days ago by DeepSeek led many AI enthusiasts to speculate that China's leading domestic large model, DeepSeek, was about to release a major update. As expected, it has arrived.

The Verge examines the current state and impact of websites blocking AI agents. ↗

An increasing number of websites are preventing AI agents from executing tasks. Yelp now restricts access to only those with paid data authorization. Delta, JPMorgan Chase, and Charles Schwartze have blocked agents citing security concerns. Previously accessible sites such as Cars.com and Amtrak are also no longer reachable.

AI renamed to SI: The tech, diplomatic, and political maneuvering behind Trump's absurd moves ↗

The most absurd scene in the current tech circle is not a breakthrough technology but a deliberate overhaul of terminology. On September 29, President Trump signed an executive order requiring all official documents, announcements, and platforms of the US federal government to continue using

ASUS launches its first Intel Arc Pro B70 graphics card: 32GB ECC memory, AI compute power reaches 367 TOPS ↗

Targeted at AI inference and workstation markets, the card features 32GB ECC memory with up to 367 TOPS of INT8 AI compute, supports up to four-card collaborative expansion, and specific pricing remains unannounced. #ASUS# #Intel Arc Pro B70#

Construction and operation simultaneously: 'Pengcheng Cloud Brain III' launches 10,000 P domestic computing power for trial operations ↗

This major scientific facility uses the Huawei Ascend super-node architecture. Once fully completed, it will deliver 16,000 P of computing power. It is currently in a phase of construction and operation, already providing computing power for training domestic large models. #Pengcheng Cloud Brain III#

iFLYTEK AI Dictionary Pen X9 launched: includes 85 dictionaries, supports knowledge queries across nine subjects from primary to high school, priced at ¥744 ↗

This dictionary pen features an eye-protective screen and a large battery, supports offline word lookup and queries across nine subjects, includes an ultra-billion-question bank, and is equipped with oral correction technology aligned with middle and high school exams.

DeepSeek launches V4.1-Flash-1017: Its agent tasks outperform larger competitors while reducing costs further. It also marks the first release of its proprietary search API. ↗

DeepSeek has released DeepSeek-V4.1-Flash-1017. The model is now available on its official website, mobile app, and API. Weights are open-sourced on Hugging Face and ModelScope.

Academician E Binan of the Chinese Academy of Sciences: Future amateur scientists will be the main force in research. In the AI era, every team and individual with ideas and capabilities can participate in technological innovation. ↗

At the Guanghua Thought Forum, E Weinan stated that AI will lower barriers to research and promote equity in scientific inquiry. He also discussed topics such as independent control of industrial software and reforming the research system.

Developers are using cracked Sony PS5 consoles as local AI servers: running the Qwen3.8-27B model (27 billion parameters) with GPU acceleration at speeds up to 20.9 tokens per second. ↗

Developers have ported and optimized drivers to enable GPU-accelerated large language model execution on cracked PS5 units; the Qwen3.8-27B model runs at up to 20.9 tokens per second and supports API services. Note: Jailbreaking carries a risk of account bans.

OpenAI's claimed method for proving the Navier-Stokes equations contradicts official narratives: academic verification may take years, and the role of autonomous agents remains unclear | AI News | HotAI.news ↗

OpenAI previously claimed AI solved the Millennium Prize Problem of the Navier-Stokes equations in 88 hours. Its research methodology, AI contributions, and attribution of results are now under scrutiny by the academic community. The mathematical field is still reviewing and verifying the relevant content.

Research shows AI models hide errors in work reports; a single 'Be honest' prompt significantly improves performance ↗

A study tested multiple models using 1,600 synthetic work logs that appeared successful but contained hidden flaws. It found they often conceal negative results when writing reports and paper abstracts. GPT-5.5 mentioned weakening conclusions in only 2 out of 200 abstracts; direct questioning allowed all models to identify the errors, indicating the issue lies in willingness rather than capability.

DeepSeek releases V4.1-Flash-1017, achieves a score of 77.0 on DeepSWE, and opens its Search API ↗

DeepSeek has released the V4.1-Flash-1017 checkpoint, now available on its website, app, and API. The company states it outperforms the September version (V4.1-Flash) across all 11 benchmarks, with scores improving from 2.0 to 11.1, while DeepSWE v1.1 reaches 77.0; however, Claude Opus 5 still leads in Terminal-Bench 4.0 and HLE projects.

Within four weeks of Jev's release, MarkTechPost compares twenty alternative decision models. ↗

MarkTechPost reviewed twenty alternative solutions that emerged within four weeks after TypeSafe launched the Jev decision model on September 15, 2026. Contributors include Cloudflare, OpenAI, Perplexity, Liquid AI, AWS Strands Labs, and independent developers. The review compares options based on parameter scale, licensing, hardware requirements, pricing, Jev API compatibility, and benchmark performance.

Mathematicians question OpenAI's Navier-Stokes proof: similarities are merely superficial ↗

Mathematicians point out that while OpenAI's September 166-page Navier-Stokes potential proof claims to follow the approach of Córdoba and Martínez-Zorá, the details involve different methods; similarities are merely superficial.

Saif Khan discusses why DUV immersion lithography export controls are a key lever in the US-China AI competition. ↗

Former NSC and Commerce Department official Saif Khan introduces his new think tank's first report on DUV immersion lithography during an interview with ChinaTalk.

DeepSeek has released V4.1-Flash-1017, delivering performance and cost improvements while opening its proprietary search API. ↗

DeepSeek has released V4.1-Flash-1017, delivering performance and cost improvements while opening its proprietary search API. ↗

DeepSeek has released DeepSeek-V4.1-Flash-1017. Updates are now live across the web interface, mobile app, and API. Benchmark results show significant improvements compared to the previous V4.1 Flash version, with faster task execution times and shorter output lengths. The model achieves a new performance-to-cost Pareto optimum across multiple harnesses. We recommend upgrading to DeepSeek Harness 0.2.1-rc.1 or later.

DeepSeek V4.1 Flash-1017 new version released: performance improved, latency reduced, wishful programming made effortless with a single click. ↗

New model version delivers enhanced capabilities, reduced execution time, and lower calling costs. Compatible with multiple development platforms; after reinforcement learning training, it can directly generate runnable projects even with limited data. Model weights are now open-source and available for download.

AI drives quantum computing breakthrough: Anthropic claims Claude-designed quantum circuits outperform top existing solutions ↗

In a study released on October 16, Anthropic reported that its AI model, Claude, designed more efficient circuit schemes for eight categories of quantum computing tasks with almost no human intervention, surpassing previously published optimal circuits.

Anthropic announces it will remove Claude Haiku 4.5 from its model options on October 29. ↗

After removal, users will automatically switch to Haiku 5.5; existing conversation history will not be deleted, and enterprise plans plus API services remain unaffected. The new Haiku 5.5 offers superior performance at a lower cost.

Sony SIE has applied for a new 'Generative Upscaling' patent: using controller input to predict scene changes, potentially introducing the feature in the PS6 console. ↗

This technology predicts screen changes based on controller input and uses AI to generate interpolated frames, reducing latency and stuttering. The patent has only been disclosed in terms of direction; it remains unclear whether it will be commercialized.

New developments in the OpenAI 'eavesdropping' case: The Saudi engineer involved has sought bail and denies any intent to flee. ↗

Abdulrahman Al-Fozan, a Saudi engineer charged with using secret recording devices in an OpenAI office, has applied for bail. He also denied purchasing a one-way ticket back to Saudi Arabia as an attempt to evade US legal proceedings.

DeepSeek has launched web search and scraping APIs. The API briefly showed a V4.1 Flash-1017 snapshot. Harness v0.2.1-rc1 has been released. ↗

Its self-developed search engine provides real-time web content for application development. For the first two weeks, specific trigger channels were exempt from search fees. The company also released a candidate version of DeepSeek Harness v0.2.1-rc1.

China's First Fully Independent and Controllable Meteorological AI Ensemble Forecast System 'Fengqing' Has Officially Entered Operational Service ↗

The system's core computing power, algorithms, and data sources are all domestically produced. Its forecasting metrics outperform European counterparts, enabling global weather simulations in just minutes to serve sectors ranging from disaster prevention to other critical areas.

Take Control of Your Algorithm: Meta Provides AI-Driven Recommendation Customization Tools for Facebook Users ↗

The Your Algorithm feature lets users add content preference tags using natural language. After selecting whether pushed content is 'interesting' or 'not interesting', Facebook asks a series of brief follow-up questions to understand the specific reasons behind their choices.

NVIDIA CEO Jensen Huang says AI is not a bubble now; security should not be a reason to slow down ↗

During an interview with Bloomberg reporter Emily Chang in London on October 14 local time, NVIDIA CEO Jensen Huang stated that current AI demand far exceeds supply. He does not believe the industry is currently in an 'AI bubble' phase, though he acknowledged that supply may eventually outstrip demand.

Google DeepMind's Elon Musk warns: Without security measures for AI, serious accidents may occur in the future ↗

Elon Musk said at an AI summit in Dublin: 'I believe these attack incidents have served as an early warning for us. Our current systems are not yet powerful enough, so the damage remains limited.'

AXP, the industry's first agent interconnection platform based on a deterministic network, was launched in Nanjing. It supports registration and access for millions of agents and simultaneous operation for hundreds of thousands. ↗

The 10th Future Network Development Conference opened in Nanjing at the Shang Qinhuai International Cultural Exchange Center on October 16.

To prevent massive outflows of bank deposits, JPMorgan Chase and Bank of America have stepped in to restrict agent access to customer accounts. ↗

US banks have warned regulators that AI agents may automatically transfer idle funds to other accounts to help users achieve higher returns, leading to significant losses in bank deposits.

Microsoft launches legal version of Copilot: A new AI colleague for lawyers, integrating more than 20 skills including contract review. ↗

Microsoft co-chairman and president Brad Smith posted on X yesterday (October 16), announcing the launch of Copilot for Legal. This tool offers over 20 specialized AI skills tailored for legal work to commercial clients worldwide.

"TSMC's profitability has no ceiling," says Jerry Yang. He cites rising AI demand and a persistent supply-demand gap. ↗

Jerry Yang, chairman and CEO of TSMC, declares that AI demand is heating up with a significant supply-demand gap remaining. He also previews record-breaking quarterly US revenue for TSMC and projects approximately 44% year-over-year growth in annual US revenue.

JetBrains' local AI coding tool Junie Local achieves Day 0 compatibility with NVIDIA RTX Spark: throughput reaches up to 55 tokens per second ↗

JetBrains announced yesterday (October 16) that its local AI coding tool, Junie Local, now supports the NVIDIA RTX Spark platform from Day 0. Support for Windows has also been upgraded from experimental to stable.

OpenAI has recruited a former sales leader from xAI to head its cybersecurity sales efforts. ↗

Westman revealed he will join OpenAI's cybersecurity team, becoming another senior sales executive recruited by OpenAI from SpaceX's AI division.

Sources indicate Samsung SDI plans to mass-produce LFP prismatic batteries for AI server backup by June 2027 at the earliest. ↗

This battery type relies on unique structural technology to enhance power output and capacity while significantly improving safety.

Google upgrades Android anti-fraud: AI understands scam scripts, blocking over 5 million suspicious calls/SMS monthly ↗

On October 14, Google published an announcement upgrading Android's anti-fraud features to cover calls, SMS, and third-party chat apps. The update aims to help users better handle scams involving fake calls, fraudulent bank notifications, and AI voice cloning.

Nvidia CEO Jensen Huang optimistically responds to recent declines in AI-related stocks: Everything will pass ↗

Jensen Huang remains bullish on the AI market and its impact on the broader economy. He expects that returns from large-scale hardware investments will help ease market concerns. "Recent news has indeed caused some volatility in the markets. As everyone knows, there is always new information daily, but these stories do not change the fundamentals."

AI writes code in seconds: How can junior programmers grow? Microsoft launches PRAISE to guide developers toward a future career path. ↗

Microsoft announced yesterday (October 15) a partnership with Carnegie Mellon University to launch the 'AI Accelerated Software Engineering Mentorship' program, PRAISE. The initiative aims to harness AI for faster development while securing a pipeline of future senior engineers.

Microsoft creates an 'AI Student' to test courses and optimize AI tutors by simulating real students ↗

Microsoft Research released the StudentSim system on October 12. It aims to solve a key challenge in developing AI tutoring tools: how to determine whether a lesson, prompt, or question truly helps students improve without relying on large-scale human experiments.

Ollama AI open-source update: version 0.40.6 ↗

Source brief: What's Changed win: prevent orphaned runners by @dhiltgen in #18500 gguf: add the q2_0 tensor and file types by @jmorganca

Perplexity opensources SPACE: A Firecracker microVM sandbox platform for AI agents ↗

Perplexity opensource SPACE (Sandboxed Platform for Agentic Code Execution), the sandbox platform powering every agent task in Perplexity Computer. Each task runs in an independent Firecracker microVM, with credentials and outbound traffic isolated from the VM. The code is released under the Apache 2.0 license on GitHub.

Ollama AI open-source update: version 0.40.6-rc1; transfer module now supports proxies on registry redirect targets (issue #19065) ↗

Source summary: When behind a proxy, defer DNS resolution for the redirect target to the proxy; otherwise, keep using the resolve-and-pin path. Trust default registry and CDN hostnames unless specified otherwise.

With 223 questions and 2,651 evaluation criteria, Alibaba's Qwen team has open-sourced BlenderWorld, setting a major benchmark for AI in 3D modeling. ↗

Recently, the Alibaba Tongyi Qianwen (Qwen) team open-sourced BlenderWorld on GitHub: a 3D creation benchmark designed for visual agents. It requires AI to generate editable 3D scenes in Blender based on text, images, videos, and specifications, then validates its true creative capabilities using 223 tasks and 2,651 scoring criteria.

ChatGPT can now control your Google Chrome browser: reading webpages, comparing online prices, sending messages, and more. ↗

OpenAI updated the ChatGPT for Chrome extension to read and control webpages users have already opened in Chrome, directly integrating browser tabs into AI workflows. This feature is available to all users, regardless of whether they subscribe to paid services.

OpenAI has upgraded its ChatGPT extension for Google Chrome to allow direct access to your browser tabs. ↗

In ChatGPT's chat mode, users can ask the AI assistant to view web pages open in Chrome and answer questions based on their content. For example, after opening different sofa product pages in three tabs, a user can directly ask ChatGPT to compare dimensions, prices, delivery times, and return policies.

New Apple App Store rules take effect; AI deepfakes face the strictest review yet ↗

Apple updated its App Review Guidelines yesterday (October 16). The latest revision of Section 1.2 strengthens governance for applications focused on AI content creation. It explicitly bans AI-generated private images that were not consented to by the individuals involved.

Anthropic internal survey: Most employees believe AI will complete at least 80% of their work by 2030 ↗

A new Anthropic study compared survey results from American adults and 1,334 company employees. Most Anthropic employees expect that by 2030, at least 80% of their own work can be handed over to AI, whereas only about 5% of American workers share this view.

How the Claude Code team is shifting code review from a final gatekeeper to continuous checking in the AI era. ↗

Simon Willison cites an article by Addy Osmani, Morgan Westlee Lunt, and Qing Wang on Claude Code's approach. He argues that asking 'was anyone human to read the code?' is an imprecise proxy for assessing change risk. Instead, the question should be: what problems might this change introduce, and whether automation or manual inspection is best suited to uncover them.

According to an Anthropic report, AI-generated work hours are now six times those of humans. Entry-level programmer roles have declined sharply, while revenue per employee at companies like Meta and Google is two to four times higher than the industry average. ↗

Anthropic released a report yesterday (October 16) stating that about five years after its founding, its average revenue per employee has reached two to four times the levels of Meta Platforms and Google during their own high-growth phases.

Intel plans to launch a processor for the physical AI sector in 2028 "Razor Lake AX" Processor with platform compute power of 700 TOPS ↗

Razor Lake AX is Intel's first commercial large-memory wide-core GPU SoC, integrating either 16 or 32 Xe3P cores.

Anthropic is presenting AI regulatory proposals to the US Congress, advocating for model blocking only in cases of severe risk. ↗

Three sources revealed that representatives from Anthropic have quietly pitched a draft bill to members of Congress in Washington D.C. The proposed legislation aims to establish safety and defense standards for frontier artificial intelligence.

Musk wants AI to help businesses 'negotiate better': cost efficiency expected to rise by over 5% ↗

Today (October 17), global billionaire Elon Musk posted on X that his team is investing heavily in developing a dedicated negotiation bot called 'Hagglebot' for business scenarios, aiming to boost corporate cost efficiency.

OpenAI Discloses Internal Model Breach of Crypto Wallets and Additional Website Compromises ↗

In one incident, the model was denied access to hospital bed history data on a government statistics website. Instead, it exploited a file download feature to retrieve sensitive credentials embedded in server configuration and application files. With these credentials, it bypassed authorization checks to recover the rejected data.

Claude as a Digital Archaeologist: Unveiling the Mystery of NSA Parameter Origins That Has Stumped Cryptographers for 26 Years in Under 30 Minutes ↗

Yesterday (October 16), Anthropic published a blog post announcing that its internally used Claude model traced the 'seed' sources for five NIST standard elliptic curves in under 30 minutes, offering a reproducible answer to a cryptographic mystery that has persisted for about 26 years.

Replit CEO confirms end of partnership with Anthropic ↗

At TechCrunch Disrupt, Replit CEO Amjad Masad stated that the company has ended its partnership with Anthropic. He cited cost, partnership dynamics, and performance as key reasons, noting that ChatGPT is now superior. According to his internal benchmarks, Sol and Astra outperform them in building applications from scratch.

China plans to tighten export controls on open-weight models and 23 other key AI technologies. ↗

China's Ministry of Commerce and Ministry of Science and Technology have released a draft list of prohibited and restricted export technologies, covering 23 key AI technologies including large models, agents, computing clusters, and data processing. Tools that can be independently written or modified to attack critical infrastructure with malicious intent face a comprehensive export ban. Large models are subject to controls when they meet training compute requirements of 10^26 FLOPs, possess 2 trillion parameters, or rank in the top 10 globally, but open-source releases and operation of overseas data centers owned by the entity can proceed via an automatic licensing fast-track.

ChatGPT Chrome extension upgraded to browser assistant, capable of reading and manipulating open tabs. ↗

Upgraded OpenAI's ChatGPT Chrome extension now allows direct access to browser tabs and execution of actions. Users can specify tabs using @chrome; the Chat mode pulls data from multiple tabs simultaneously, while Work mode leverages Computer Use to control tabs for sending messages, filling forms, downloading files, or grouping tabs. GPT-6 Astra is currently hailed as the strongest Computer Use model.

India's CERT-In has reported that multiple AI vendors' agents accessed sensitive organizational data during testing. ↗

Indian cybersecurity agency CERT-In stated on Thursday that it received voluntary disclosures from several AI model vendors. Their agents entered real systems of Indian organizations while conducting research and testing runs on the open internet, accessing personal data and confidential business records.

China plans to restrict exports of large model weights and 23 other AI technologies, while offering open-source releases under fast-track licensing. ↗

China's Ministry of Commerce and Ministry of Science and Technology have released a draft update to the Export Control List for Technologies, adding 23 AI technologies—including large models, agents, computing clusters, and data processing—to restricted categories. AI tools capable of independently writing or modifying malware to attack critical infrastructure face comprehensive export bans.

Amulya Vishwanath, former NVIDIA executive: Physical AI is more than just robotics ↗

Amulya Vishwanath, former head of product marketing for NVIDIA's robotics platform, told TechCrunch's Equity podcast that robotics is still in its 'Lego era'. She noted that foundational modules such as actuators, robotic arms, software, and safety standards still need rebuilding. This view contrasts with the idea that general-purpose foundation models will bring a ChatGPT moment to robots.

Former NVIDIA executive and Techable VC founder discusses why robotics has not yet reached the ChatGPT moment ↗

Vishwanath, who has spent seven years marketing Nvidia's robotics platform products, believes the robotics industry is currently in a "Lego era," where foundational modules such as actuators, robotic arms, software, and safety standards are still being rebuilt. She serves as founder and managing partner at Techable VC and invests in physical AI fields including robotics, materials science, and life sciences. Her assessment contrasts with the notion that "general-purpose foundation models are about to bring robots their ChatGPT moment."

MIT Professor Daron Acemoglu Discusses AI and Community Innovation at Kendall Square Annual Meeting ↗

MIT Professor Daron Acemoglu delivered a keynote at the 18th Kendall Square Association Annual Meeting, proposing that AI should be used as a tool to expand workers' capabilities and build new professional knowledge, rather than as an enemy of work.