Local timeline

Anthropic's Claude is now available in India via Amazon Bedrock, offering in-country inference. The service includes audit trails and access controls for risk and compliance teams.
Source: BusinessLine
https://www.thehindubusinessline.com/info-tech/anthropics-claude-live-in-india-with-in-country-inference-through-amazon-bedrock/article71545810.ece
#EnterpriseAI #Anthropic #Claude

Japanese contract chipmaker Rapidus plans to partner with 17 domestic and international design firms to broaden its potential customer base and help clients develop proprietary silicon.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261005VL205/rapidus-silicon-design-manufacturing-automation.html
#Semiconductors

The US Trump administration has appointed a new AI czar and task force. They have 120 days to report on AI's risks and opportunities and recommend the federal government's role.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261005VL207/competition-government-technology.html
#TechPolicy

Broadcom has agreed to provide up to US$42 billion in convertible-note financing to support Anthropic's infrastructure spending.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261005PD220/financing-anthropic-broadcom-ipo-supplier.html
#Semiconductors #Broadcom #Anthropic

SafeStreamingFlow aligns flow sampling dynamics with execution dynamics. arXiv cs.RO
Source: arXiv cs.RO
https://arxiv.org/abs/2610.03132
#Robotics

SocialVLA converts spontaneous human reactions into runtime intervention signals for VLA manipulation.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.02360
#Robotics

Skill2Real is an agentic policy framework that learns executable robot skills through a shared application programming interface.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.02788
#Robotics

A learned proprioceptive reflex policy maps task-space commands to joint-position targets using state history.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.02811
#Robotics

Filter-Aware Fine-Tuning for Safe Humanoid Whole-Body Tracking presents CoFiT, a method for pretrained trackers.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.02341
#Robotics

Researchers propose GeoScaffold to internalize geometry into vision-language navigation policies at training time.
Source: arXiv cs.RO
https://arxiv.org/abs/2610.02697
#MachineLearning

Researchers introduced Lexicographic Multi-Objective On-Policy Distillation to integrate reward-specialized policies under explicit priorities.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02359
#MachineLearning

ArrivalBench evaluates agent-generated data pipelines under replayable delivery schedules instead of a fixed snapshot. Re-executing pipelines finds that many certified pipelines are silently wrong.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02363
#MachineLearning

Researchers study long context language models by training a model over a simple harness with tools to call itself and read tokens.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02404
#MachineLearning

MuLoRA introduces spectrally balanced low-rank adaptation to mitigate spectral plasticity collapse in continual learning.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02283
#MachineLearning

PowerBench measures language model bias in power-shifting requests, distinguishing self-empowerment, disempowerment, and power grabbing. Models refuse power grabbing more than disempowerment.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02303
#MachineLearning

Researchers developed a stability theory for Joint-Embedding Predictive Architectures that unifies heuristics.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02344
#MachineLearning

The AI Theorist system autonomously generated a new interpretation of optical and photocurrent data for \alpha-RuCl$_3$.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02417
#MachineLearning

Researchers introduced Co-Design Gym, a suite of benchmark environments for jointly optimizing embodiment and policy.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02366
#MachineLearning

LLM agents rely on activation probes as runtime monitors to catch harmful inputs. A classification instruction appended after the user's turn can improve out-of-distribution detection.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02413
#MachineLearning

A new study investigates why adaptive batching helps large language model pretraining.
Source: arXiv cs.LG
https://arxiv.org/abs/2610.02355
#MachineLearning