Timeline lokal

OpenAI has filed a lawsuit claiming a former employee who joined Apple took files concerning an unreleased device.
Source: Gadgets 360
https://www.gadgets360.com/ai/news/openai-lawsuit-employee-apple-files-unreleased-device-12193868#rss-gadgets-all
#AI #OpenAI #Apple

Tencent stated that AI agents used to evaluate third-party models displayed web-browsing behavior not intended by the tech giant. The review found no evidence of deliberate targeting or coordination.
Source: BusinessWorld
https://bworldonline.com/technology/2026/10/16/787441/chinas-tencent-says-ai-agents-showed-unintended-browsing-behaviour/
#AIAgents #Tencent

AgiBot is rolling out its YuanSheng ecosystem to more than 20 Chinese universities, planning to place over 1,000 robots on campus.
Source: Pandaily
https://pandaily.com/agibot-yuansheng-university-program-20-schools-1000-robots-aima
#Robotics

WanNa Robotics introduced NOVA21, a 21-degree-of-freedom dexterous hand aimed at research labs and robot developers.
Source: Pandaily
https://pandaily.com/wanna-robotics-nova21-21-dof-direct-drive-dexterous-hand-hong-kong
#Robotics

Samsung and SK Hynix are developing new memory technologies to address growing bottlenecks in AI systems. Samsung is exploring a design that stacks high-bandwidth memory directly above graphics processors, while SK Hynix is advancing high-bandwidth flash.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261016PD214/samsung-sk-hynix-hbm-data-capacity.html
#Semiconductors #Samsung

Axera unveiled the AX8860 edge AI chip series with up to 256 TOPS and 128 GB of memory.
Source: Pandaily
https://pandaily.com/axera-ax8860-edge-ai-chip-256-tops-128gb-lpddr5x-axlink-ciif-2026
#Semiconductors

India's cyber agency says AI agents breached systems.
Source: The Hindu
https://www.thehindu.com/sci-tech/technology/after-openais-australia-hack-indias-cyber-agency-says-ai-agents-breachedsystems/article71591537.ece
#TechPolicy #OpenAI

Hesai launched Kosmo, a 3D world data engine designed to transform physical environments into high-fidelity 3D digital spaces.
Source: TechNode
https://technode.com/2026/10/16/teaching-robots-the-shape-of-reality-hesai-s-kosmo-turns-real-places-into-3d-assets/
#Robotics

EvolvingBrain introduces an adaptive skill system designed for multi-modal agent self-evolution in open worlds.
Source: arXiv
https://arxiv.org/abs/2610.20786
#MachineLearning

Modality Self-Play is a training framework for multimodal foundation models based on proxy-mediated modality composition.
Source: arXiv
https://arxiv.org/abs/2610.20965
#MachineLearning

OmicsBench is a long-horizon, multi-turn bioinformatics benchmark for AI agents. It links 97 tasks into a directed acyclic graph covering 662 subtasks.
Source: arXiv
https://arxiv.org/abs/2610.21074
#MachineLearning

Kolab is a new dataset of 115+ hours of screen recordings of 50 professionals from 40 occupations collaborating with agents.
Source: arXiv
https://arxiv.org/abs/2610.21109
#MachineLearning

LLM-based planning agents commit to plans of different lengths under partial observability. Aggressive and conservative planners reach similar accuracy across three LLMs.
Source: arXiv
https://arxiv.org/abs/2610.21108
#MachineLearning

Anchor has launched Action Index, a shared memory of web workflows that agents have completed successfully.
Source: The Next Web
https://thenextweb.com/news/anchor-action-index-computer-use-agents-cost
#AIAgents

All-in-One Memory proposes a process-rewarded reinforcement learning framework that treats retrieval, writing, and stopping as actions evaluated in a shared reward space.
Source: arXiv
https://arxiv.org/abs/2610.21147
#MachineLearning

Benchmarks lose their differentiating power and AI safety utility when many models achieve perfect scores.
Source: arXiv
https://arxiv.org/abs/2610.20667
#MachineLearning

MAI-Gym is an open platform that unifies GUI agent evaluation, data collection, and online reinforcement learning training across mobile, computer, and web environments.
Source: arXiv
https://arxiv.org/abs/2610.20939
#MachineLearning

A self-improving agent can grow its own domain-specific language without reward, a judge model, or answer keys and still produce exactly correct answers.
Source: arXiv
https://arxiv.org/abs/2610.20809
#MachineLearning #Oracle

StepSafe introduces a training framework for step-level timely intervention in agentic systems, aligning decisions with prefix stages and using adaptive feedback self-distillation.
Source: arXiv
https://arxiv.org/abs/2610.20960
#MachineLearning

Microsoft will scale its Maia accelerator family to multi-gigawatt deployments from 2027, making it its primary inference system.
Source: InfoQ
https://www.infoq.com/news/2026/10/azure-maia-300/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global
#Semiconductors #Microsoft