Timeline lokal

Current and former OpenAI and Google DeepMind researchers say companies are doing too little to protect the world from self-improving AI systems that could outpace human control.
Source: Rappler
https://www.rappler.com/technology/features/ai-researchers-warn-companies-rush-self-improving-systems-risks/
#TechPolicy #OpenAI #Google

Venture capital investment in European defencetech startups has reached a record $7.4 billion in 2026, nearly three times the amount raised in 2025.
Source: Tech.eu
https://tech.eu/2026/10/05/european-defencetech-hits-record-7-4b-as-investment-nearly-triples-in-2026/
#Startups

Researchers introduced FD-SCoPE, a language-model framework answering clinician questions over trial evidence tables.
Source: arXiv
https://arxiv.org/abs/2610.02576
#MachineLearning

MintFlow is a training-free constrained sampling framework that enforces constraints by minimally perturbing an intermediate flow state while keeping the pretrained flow field unchanged.
Source: arXiv
https://arxiv.org/abs/2610.02260
#MachineLearning

Large language models spend the same amount of computation on every token they generate.
Source: arXiv
https://arxiv.org/abs/2610.02491
#MachineLearning

Large language models rely on long-horizon tool invocation sequences for complex tasks. CITA trains a Comparative Inference Model from paired signals to estimate tool invocation values.
Source: arXiv
https://arxiv.org/abs/2610.02330
#MachineLearning

MEA uses a Proposer and an Actor agent to generate faithful natural-language explanations across modalities, yielding gains over an untrained backbone.
Source: arXiv
https://arxiv.org/abs/2610.02480
#MachineLearning

Apple announced it will introduce additional controls in macOS for apps seeking Full Disk Access. The company named AI agents as a key reason.
Source: MediaNama
https://www.medianama.com/2026/10/223-apple-consent-mac-muse-dispute/
#TechPolicy #Apple #Meta

Open-Endedness Bench evaluates the epistemic process of agents on open-ended research tasks by reading execution records.
Source: arXiv
https://arxiv.org/abs/2610.02588
#MachineLearning

DuplexJev feeds ASR-encoder hidden states into a frozen LLM to read each question as a single-token distribution.
Source: arXiv
https://arxiv.org/abs/2610.02638
#MachineLearning

Using Claude Opus as a meta-agent to generate terminal tasks and verifiers for reinforcement learning training is becoming common.
Source: arXiv
https://arxiv.org/abs/2610.02405
#MachineLearning #Claude

Researchers introduced VERSE, a verified self-evolving optimizer for agent harnesses that improves prompts, tools, and workflows while model weights stay fixed.
Source: arXiv
https://arxiv.org/abs/2610.02616
#MachineLearning

A new protocol for AI debate improves on prior work by ensuring correctness in a worst-case sense and achieving instance-wise optimality.
Source: arXiv
https://arxiv.org/abs/2610.02557
#MachineLearning

President Donald Trump announced a Super Intelligence Force to coordinate federal work on advanced AI, reporting to him and White House Chief of Staff Susie Wiles.
Source: Techxmedia
https://techxmedia.com/en/us-super-intelligence-policy-takes-shape-after-ai-rename/?utm_source=rss&utm_medium=rss&utm_campaign=us-super-intelligence-policy-takes-shape-after-ai-rename
#TechPolicy #WhiteHouse

HXAI is a hierarchical framework that preserves privacy while enabling explainable analysis for grid-level demand management.
Source: arXiv
https://arxiv.org/abs/2610.02504
#MachineLearning

The Dubai Electronic Security Center has unveiled SARAAB, an artificial intelligence model designed to detect deepfake videos in seven second clips. SARAAB achieved up to 91 per cent detection accuracy and 89 per cent accuracy on previously unseen data.
Source: Gulf News
https://gulfnews.com/uae/new-dubai-ai-model-spots-deepfakes-in-seven-second-clips-1.500698366
#AI

A new paired and self-audited evaluation investigates System-1 decision models for LLM agent harnesses, comparing an open-weight model and a hosted model across eleven agent decision points.
Source: arXiv
https://arxiv.org/abs/2610.02267
#MachineLearning

Researchers propose CuBEs to evaluate large language model behaviors across diverse user cultures. Cultural context changes behavioral evaluation outcomes.
Source: arXiv
https://arxiv.org/abs/2610.02622
#MachineLearning

AI-powered cyberattacks have targeted Korea's financial institutions, showing the nation is ill-prepared for the new age of artificial intelligence.
Source: The Korea Times
https://www.koreatimes.co.kr/opinion/editorial/20261005/ed-koreas-ai-blind-spot?utm_source=rss
#TechPolicy

Huawei agreed to a multi-year patent licensing deal with Qualcomm covering artificial intelligence, 5G, computing and networking technologies.
Source: The Economic Times
https://economictimes.indiatimes.com/tech/technology/huawei-agrees-to-a-multi-year-patent-licensing-deal-with-qualcomm/articleshow/134687782.cms
#AI #Huawei #Qualcomm