Local timeline

Taiwan's semiconductor supply chain is rushing to Texas following reports of a collaboration with Elon Musk's "Terafab" plan. DYC chairman Chung-Ping Hsieh said his company received the news last week and is mapping out a Texas strategy.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261006PD236/tsmc-texas-supply-chain-taiwan.html
#Semiconductors #TSMC

China urged the United States to adopt a more open approach and strengthen AI cooperation with developing countries, including China. US President Donald Trump announced a new federal task force named the Super Intelligence Force to coordinate government efforts.
Source: Phnom Penh Post
https://phnompenhpost.com/international/china-urges-us-to-collaborate-more-in-the-field-of-ai/
#TechPolicy

Autonomous coding agents solve repository issues through various actions. Extra inference-time compute yields gains only when producing a useful repair and reliable evidence.
Source: arXiv
https://arxiv.org/abs/2610.03984
#MachineLearning

SkillScriptBench evaluates executable agent skill self-evolution. We introduce AST-Guided Skill Revision.
Source: arXiv
https://arxiv.org/abs/2610.04008
#MachineLearning

Sparse Autoencoders decompose model activations into sparse combinations of interpretable dictionary atoms. Their objective smuggles in an additional prior where concepts across patches are treated as independent. Spatial-SAE is proposed as an amortized MAP estimator.
Source: arXiv
https://arxiv.org/abs/2610.04112
#MachineLearning

The framework combines a neural network for routine driving with model predictive control for unfamiliar situations. Tests in CARLA show it reduces collisions by 89%.
Source: arXiv
https://arxiv.org/abs/2610.04088
#MachineLearning

NLIP and A2A protocols are benchmarked across three hardware environments.
Source: arXiv
https://arxiv.org/abs/2610.04053
#MachineLearning

LatentQuant is a two-stage NVFP4 QAT framework designed to preserve the policy-facing latent contract under VAE quantization.
Source: arXiv
https://arxiv.org/abs/2610.03959
#MachineLearning

Researchers measured how well one expert human reader and three LLM families reproduced their own reading of Sumerian myths over time. Across runs months apart, the human cut the text in much the same places but named the segments differently.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04370
#MachineLearning

The Australian Broadcasting Corporation told a parliamentary inquiry that AI companies have probably already used its content to train their models.
Source: The Next Web
https://thenextweb.com/news/abc-tells-australian-ai-inquiry-its-content-has-likely-been-scraped
#TechPolicy

Cybersecurity training company Hack The Box Ltd. introduced AI Range Enterprise Edition to let security teams test whether their artificial intelligence agents can do their assigned jobs.
Source: SiliconANGLE
https://siliconangle.com/2026/10/06/hack-the-boxs-ai-range-enterprise-edition-tests-whether-ai-agents-can-do-their-jobs/
#AIAgents

CADForge is an agentic framework proposed to reconstruct editable parametric CAD models from a single-view image by converting images into CadQuery programs.
Source: arXiv
https://arxiv.org/abs/2610.04262
#MachineLearning

AMD plans to increase its investment in Taiwan's semiconductor supply chain as demand for CPUs, GPUs and AI computing exceeds capacity.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261006VL219/amd-capacity-investment-demand-taiwan-supply-chain.html
#Semiconductors #AMD

Researchers question similarity scores used to claim shared structure between the brain or languages. A probe performance drops to chance in four of seventeen languages, and a model trained on a destroyed brain target still scores 0.31.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.03827
#MachineLearning

Moving from zero to one critique round produces 34.5 percentage points of additional mechanism-level divergence.
Source: arXiv
https://arxiv.org/abs/2610.04133
#MachineLearning

Language model activations inhabit privileged error-correcting basins. Researchers hypothesize this robustness results from passive dynamics constraining mechanisms in the forward pass.
Source: arXiv
https://arxiv.org/abs/2610.04183
#MachineLearning

The paper defines agentic cognitive depth as a trajectory-level profile across five operational criteria.
Source: arXiv
https://arxiv.org/abs/2610.04168
#MachineLearning

Studies examined generative Agentic AI for spatial rotation.
Source: arXiv
https://arxiv.org/abs/2610.04188
#MachineLearning

Language-model systems can separate contextual computation, persistent storage, and exact execution instead of updating all capabilities through one shared parameter system.
Source: arXiv
https://arxiv.org/abs/2610.04012
#MachineLearning

SGAnalog introduces an end-to-end circuit benchmark from open-source silicon tapeouts. It contains 273 topologically distinct top-level designs. Across seven models on 66 transcription tasks, the strongest model reaches 56.1 percent exact graph isomorphism.
Source: arXiv
https://arxiv.org/abs/2610.03934
#MachineLearning