At the Android Dev Summit, Google launched Android Studio Rabbit 2, enabling the 'Bring Your Own Agent' (BYOA) feature by opening the Agent Client Protocol (ACP) to support over 40 coding agents including Claude, Codex, and Google Antigravity. It released the Android CLI 2.0 plugin, adding terminal Compose previews, direct device debugging, and LeakCanary integration; internal tests showed a 77% speedup for UI tasks and reduced token usage to one-third. Google also introduced the ML Kit Agentic API (alpha), AppFunctions API (beta), and Agent-to-UI (A2UI) interface based on Gemini Nano 4. Additionally, it launched a preview of the Connected Apps Platform and partnered with XREAL to release the AURA smart glasses running Android XR, priced at $1099.
Anthropic has launched a new 'artifacts' feature for Claude for Legal that supports automatic verification of cited sources with links to original materials, followed by lawyer review. It introduces interactive overviews and tables for verifying citations, cross-jurisdictional legal comparisons, case file management, legislative tracking, and contract compliance reviews. The feature is integrated with more than 60 legal databases, including Bloomberg Law, and supports integration with proprietary tools from firms such as Goodwin, Harvey, Foley & Lardner, and PwC. Built-in settings enforce data protection rules and client-mandated team information segregation.
a16z released 'The New Security Agenda' report, noting that AI has become a core variable in cybersecurity ahead of quantum computing. Incidents such as OpenAI agents breaching Hugging Face production systems, AI aiding the discovery of a PHP vulnerability unpatched for 10 years, and eMudhra certificate authority using weak keys indicate that AI is accelerating the exposure and exploitation of system weaknesses.
Shopify Analytics has launched a bubble map feature that aggregates nearby shipping or billing addresses into bubbles scaled by metrics such as order volume or total sales; it supports zooming, panning, and clicking to view details. This feature is integrated into all Analytics reports and dashboard cards, and can be enabled via the 'Order Hotspots by Shipping Location' report or by adding shipping/billing address fields in any report.
SimGym now supports launching simulations with specific shopping goals, including full purchase flow, product discovery, and add-to-cart confidence; recommendations are now marked with numbers and linked directly to page positions, with each recommendation linking to an AI shopper visit summary and replay video to help identify friction points and positive response points.
Shopify Campaign Autopilot is now open to all merchants, supporting Meta Ads, Microsoft Advertising, Snapchat, Shop, Shopify Messaging, and X Ads (coming soon); it automatically analyzes stores, products, and customers to build marketing campaigns across channels and dynamically allocate budgets; it also provides reviewable and approvable SEO optimization suggestions, or allows direct requests for SEO updates via Sidekick.
Nvidia released the open-source speech model Nemotron 3.5 Transcribe, which supports transcription in 40 languages, translation between 31 languages and English, speaker identification for 13 languages, and audio-based question answering; the model has 30 billion parameters but activates only about 4 billion per step, enabling it to run on a single Nvidia GPU; it ranks first on the Hugging Face English Open ASR leaderboard with an average word error rate of 5.33%, and leads in speaker identification performance on the VoiceArena and David AI benchmarks; simultaneously, the Granary v2 dataset (844,000 hours, 48 languages) and fine-tuning tools were open-sourced; the model does not support real-time processing, and the audio question-answering feature was trained only on English.
In 2026, as public corpora dried up and training demands for ultra-large models surged, mainstream Chinese AI companies densely adjusted their data strategies and leadership: ByteDance established a company-level first-tier department named 'AI Data and Security', with former TikTok executive Wang Yinglei succeeding Fu Yue; Seed internally created a new Pretrain Data department, managed by Li Chenggang overseeing full-modality pretraining data; Tencent's Chief AI Scientist Yao Shunyu directly serves concurrently as head of the AI Data department; Alibaba showcased four specialized leaders for training data, expert ecosystem, evaluation, and embodied data at the Yunqi Conference; Xiaohongshu hired former DeepSeek core member Chen Guanting as head of pretraining data; Meituan, Moonshot AI, Zhipu, and DeepSeek also clarified their data leaders or core roles.
Based on personal practice, Simon Willison pointed out that Claude Opus 5.5 demonstrated significantly improved game design capabilities in October 2026, capable of generating playable, interesting, and aesthetically pleasing HTML/JS browser games in one go; he used this model to complete four games: Tree Appreciator (a tree-comparison casual game), Emberwick (a top-down fantasy RPG with story and terrain), Battle Drones (an RTS supporting multiplayer online play and LLM battles), and Pumpkin Farmer (a semi-fictional pumpkin farm simulation), all requiring only minimal prompts and no code intervention; he compared GPT-6 Astra's performance as slightly inferior, emphasizing that Opus 5.5's understanding of 'fun' has achieved a substantive breakthrough.
Coinbase stated that its systems were not compromised, but a security incident involving third-party data analytics platform Amplitude led to the leakage of contact information for approximately 70,000 Coinbase users. Amplitude has not yet responded; it is a SaaS analytics provider with clients including crypto institutions such as Ramp Network and Xapo Bank. Coinbase previously experienced data breaches in 2025 and 2021 caused by customer service bribery and SMS verification vulnerabilities, respectively. The article cites views from CZ and reports from Microsoft, CrowdStrike, and IBM, noting that AI is accelerating the pace of social engineering, phishing, and vulnerability exploitation, but has not introduced entirely new types of attacks.
James Comer (R-KY), Chair of the U.S. House Oversight Committee, has launched a formal investigation into the proposed Delaware AI Company (AIC) bill, requiring Governor Matt Meyer to submit all relevant documents and communications from January 2025 onward, including internal AI chat logs, by November 9; the proposal was put forward by Delaware Secretary of State Charuni Patibanda-Sanchez and Norm Ai CEO John Nay in a July Fortune column, aiming to establish an AI-driven corporate entity subject to a 30-month regulatory sandbox, granting it litigation capacity while limiting liability boundaries.
On October 23, 2026, the General Office of the State Council issued "Opinions on Adapting to the Development of Artificial Intelligence Technology to Help Build an Employment-Friendly Development Model," proposing 18 tasks to mitigate AI-driven job displacement, establish risk assessment mechanisms, and adjust income distribution and social security systems; combining corporate data such as Salesforce customer service team reductions, Klarna's AI handling 69% of customer chats and saving $39 million, and Stanford University research showing a relative employment gap of 19% for workers aged 22 to 25, the article points out that AI is reducing hiring for junior positions, weakening vocational training functions, and decoupling labor income share from productivity growth.
llama.cpp v0.8.0 has been officially released, adding model families such as LiquidAI LFM3-1.7B (text), LFM3-VL-2B (vision), LFM3-Audio-2B (speech recognition and synthesis), and LFM3-Omni-2B (multimodal speech dialogue). It also introduces an experimental streaming-audio API (mtmd_audio_stream) and /v1/realtime and /tts endpoints compatible with the OpenAI Realtime protocol. Additionally, it supports new architectures including DeepSeek V4.1 Flash, SenseNova U1, Beam, Kolibri-1, Flint, Muse Spark, and Granite Switch Decision. Features optimized include sparse attention (QSA/DSA), MoE caching (50/50 SLRU), memory management (no_alloc), direct download of HF Hub models, conversation search, and compression. ggml has been upgraded to v0.28.0, enhancing backend performance for CUDA/SYCL/OpenCL/Hexagon/Vulkan/WebGPU.
On October 26, Microsoft launched a voluntary separation program covering approximately 60,000 employees at grade level 70 and below in select teams across the United States, Canada, the United Kingdom, and India. Compensation is estimated at 16 to 20 weeks of salary, and participants are barred from rejoining for two years. This initiative follows multiple layoffs in 2025 and a voluntary retirement plan in April 2026 as another round of workforce adjustments. During the same period, the company's fiscal year 2026 capital expenditure reached $145.3 billion, Azure annual revenue exceeded $100 billion for the first time, and paid seats for Microsoft 365 Copilot surpassed 30 million.
Cybersecurity startup Cathedral used an AI agent to launch a penetration attack through third-party business software (external administrator and sales programs) used by Anthropic, claiming to have obtained meeting records, payroll data, job postings, and Slack messages in less than twelve hours; Anthropic confirmed limited access to some sales, HR, and IT systems, including partial customer and partner contact information and basic account data, but emphasized that Claude chat logs, source code, and model weights were not leaked, and stated it closed the entry point, disabled relevant third-party programs, and notified US federal authorities on the same day; there are discrepancies between the two parties regarding the duration and scope of the attack, with Anthropic stating the attack lasted several days and that Cathedral violated bug bounty program rules, while Cathedral insisted it followed standard security research practices; former OpenAI employee Will Depue questioned Anthropic's downplaying of the incident, claiming Cathedral had obtained full Slack admin privileges, while Anthropic red team member Logan Graham denied this claim but acknowledged the industry needs to restructure its security systems annually; Cognition Chief Security Officer Alex Stamos pointed out that third-party integrated applications represent a common risk across the entire industry.
LimX Dynamics' open-source embodied AI engineering foundation, FluxVLA Engine, completed a significant update on October 22. It provides unified model integration, three-step real-robot deployment (inference acceleration, RTC timing alignment, trajectory post-processing), online reinforcement learning post-training integration, and closed-loop evaluation capabilities across 14 strategies and 4 types of real-robot platforms. The technical report has been published on arXiv, and the code repository is hosted on GitHub.
Moore Threads has open-sourced its large model training infrastructure project, Megatron-MUSA-Titan. Based on its full-function GPU, the project focuses on training trillion-parameter MoE models on ten-thousand-card-level clusters, covering two major scenarios: pre-training and reinforcement learning. It supports blockwise FP8 training, DeepEP, and the ACE asynchronous communication engine, and can collaborate with slime, miles, and SGLang. The project has been continuously used in Moore Threads' ten-thousand-card-level cluster for trillion-parameter MoE pre-training, achieving high MFU and long-term stable training. The code is hosted on GitHub.
The article analyzes how surging computing power and bandwidth demands for AI chips have led to continuously expanding package sizes, approaching reticle limits; TSMC's CoWoS-L achieves a 5.5x reticle size package through RDL + Local Silicon Interconnect (LSI) and has entered mass production, expected to maintain mainstream status until before 2028; Intel's EMIB-T introduces TSVs on top of local silicon bridges to enhance power delivery capability, supporting ultra-large-scale integration, with Google and AWS currently testing this solution; both represent different 2.5D paths but jointly point toward a new stage where packaging must be co-designed with chips.
Liquid AI released the LFM3 model family, based on a new tri-mixed layer architecture, pre-trained on 22T tokens; the first batch includes LFM-1.7B (text), LFM3-VL-2B (vision), d1-2B (decision-making), LFM3-Audio (audio), LFM3-Omni-2B (omni-modal), and d1-Omni-2B (omni-modal decision-making), all based on LFM3-1.7B-Base; LFM-1.7B features a 256K context window, with 35% fewer parameters while outperforming LFM2.5-2.6B.