Timeline lokal

Researchers introduced SCOUT, a two-stage agentic safety verifier for computer-use agents that synergizes reasoning-intensive rubric generation with tool-intensive evidence gathering.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.36201
#MachineLearning

A new study evaluates Large Language Models including Llama-3 8B, Qwen-2.5 14B, and Llama-3.3 70B for extracting specific attributes from SEC 10-K text.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35864
#MachineLearning #SEC #Llama

TypeSafe AI has released Jev, a decision-only model that returns typed probabilities instead of text.
Source: InfoQ
https://www.infoq.com/news/2026/10/typesafe-ai-jev-released/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global
#Tech

Google is withholding its Gemini 4 Argon artificial intelligence model from the public, releasing it only to a vetted group of cybersecurity experts.
Source: Malay Mail
https://www.malaymail.com/news/tech-gadgets/2026/10/01/google-holds-back-gemini-4-argon-limits-release-to-vetted-cybersecurity-experts/237250
#AI #Google #Gemini

California enacted new laws targeting companies using artificial intelligence to initiate job cuts or monitor employees.
Source: Engadget
https://www.engadget.com/2274224/california-law-bans-companies-from-relying-on-ai-to-fire-workers/
#TechPolicy

Google unveiled Gemini 4 Argon, which supports a one million token limit for complex AI tasks.
Source: TechEconomy Nigeria
https://techeconomy.ng/google-gemini-4-argon-1-million-token-limit
#AI #Google #Gemini

PADM\'E is a data synthesis method that generates reliable criterion-based meta-evaluation data for agentic settings using only small language models.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.36086
#MachineLearning #Meta

Google launched Gemini 4 Argon with guardrail-free access for vetted defenders. The company says its new frontier AI model found a critical vulnerability in software used by hospitals worldwide.
Source: SecurityWeek
https://www.securityweek.com/google-launches-gemini-4-argon-with-guardrail-free-access-for-vetted-defenders/
#AI #Google #Gemini

PrimeSeeker is a capability-oriented framework that builds web-grounded anchor structures and generates both a question and a reference evidence skeleton.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35816
#MachineLearning

Researchers used causal mediation analysis to study sycophantic agreement in language models. They found that a sparse set of early attention heads carries the opinion signal and that ablating these heads substantially reduces sycophancy.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35822
#MachineLearning

Huawei rotating chairman Eric Xu said the company's Ascend AI accelerators have surpassed Nvidia in sales and market share in China.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261001PD232/huawei-ascend-ai-chip-china-nvidia.html
#Semiconductors #Nvidia #Huawei

Google withheld its most powerful artificial intelligence model, Gemini 4 Argon, from public release on October 1, 2026, limiting access to vetted cybersecurity experts and the US government.
Source: TechJuice
https://www.techjuice.pk/google-gemini-4-argon-restricted-access-safety-cyberattack-concerns/
#AI #Google #Gemini

The paper introduces Memory Adaptation for Task-Conditioned Execution, a deterministic post-retrieval procedure that converts trajectories into execution-oriented memory.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35808
#MachineLearning

Intrinsic self-correction allows a language model to revise its own answer without external evidence. A second pass recovers mistakes but can overturn correct answers. Aggregate accuracy can conceal differing revision behavior.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35832
#MachineLearning

The authors identify an over-uniform training objective as the main obstacle to scaling uniform diffusion language models.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35817
#MachineLearning

AI godfather Yann LeCun has zero concerns about human extinction and says Anthropic CEO Dario Amodei is deluded.
Source: Fortune
https://fortune.com/2026/10/01/ai-godfather-yann-lecun-has-zero-concerns-about-human-extinction-says-anthropic-ceo-dario-amodei-is-deuded/
#TechPolicy #Anthropic

Demographic Pluralism is introduced as an inference-time framework that estimates population-level opinion distributions without opinion-distribution training data or task-specific fine-tuning.
Source: arXiv
https://arxiv.org/abs/2609.38555
#MachineLearning

Visa and Mastercard handle nearly half of Europe's card payments. The digital euro aims to change that. Gilles Boyer stated that payment systems are instruments of power and noted European dependence on the U.S.
Source: Fortune
https://fortune.com/2026/10/01/europe-digital-euro-challenge-us-visa-mastercard-payments-dominance/
#TechPolicy

GoldiMask selects context tokens by approximately maximizing a submodular objective and weights remaining targets based on their learning potential.
Source: arXiv
https://arxiv.org/abs/2609.38385
#MachineLearning

Horizon Robotics released version 2.1 of HSD, adding end-to-end reversing capability. Chery's iCAR V27 is the first model to receive updates starting October 8.
Source: Pandaily
https://pandaily.com/horizon-robotics-hsd-v2-1-end-to-end-reversing-icar-v27-ota
#Robotics