Timeline lokal

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

Large language models can solve tasks whose answers depend on relations among input tokens.
Source: arXiv cs.CL
https://arxiv.org/abs/2609.35970
#MachineLearning

Aligned training can induce misaligned behavior in other contexts. This phenomenon is called context confusion and is demonstrated across three domains.
Source: arXiv
https://arxiv.org/abs/2609.38379
#MachineLearning

DIG Ventures closed its third fund at $120 million to invest in early-stage European AI infrastructure startups. The firm plans to back approximately 30 companies at pre-seed and seed stages.
Source: Tech.eu
https://tech.eu/2026/10/01/dig-ventures-closes-120m-fund-iii-to-back-europes-ai-infrastructure-startups/
#Startups

Oracle announced Fusion Claw, a governed agentic execution runtime for Oracle Fusion Agentic Applications. The runtime combines AI reasoning with deterministic enterprise computation.
Source: Techxmedia
https://techxmedia.com/en/oracle-launches-fusion-claw-for-complex-enterprise-work/?utm_source=rss&utm_medium=rss&utm_campaign=oracle-launches-fusion-claw-for-complex-enterprise-work
#EnterpriseAI #Oracle

Google unleashed AI Overviews across branded search territory in September 2026. Testing by Chris Long revealed the feature appearing on 93% of major brands tested. DemandSphere data confirms the aggressive deployment.
Source: TechJuice
https://www.techjuice.pk/google-ai-overviews-brand-traffic-displacement/
#AI #Google

Meta wants users to rely on its new Muse AI agent to handle everyday tasks such as booking flights, ordering groceries and calling customer service.
Source: Stuff South Africa
https://stuff.co.za/2026/10/01/meta-wants-people-to-use-ai-agents-to-do-tasks/
#AIAgents #Meta

OpenAI published a misalignment reports site that documents nine disclosed cases showing a pattern of lost control over its AI agents.
Source: TechJuice
https://www.techjuice.pk/openai-rogue-ai-incidents-misalignment-reports-thousands-revealed/
#AIAgents #OpenAI

Huawei Digital Power has launched a grid-interactive AIDC 1.0 solution for AI data centers.
Source: Pandaily
https://pandaily.com/huawei-grid-interactive-aidc-1-0-power-liquid-cooling-ascend-sensetime-lingang
#AI #Huawei

XPeng and Banma Intelligence have built an automotive-grade AI cockpit solution that runs a 30B-class large model entirely inside the car.
Source: Pandaily
https://pandaily.com/xpeng-banma-turing-chip-30b-autoomni-on-device-ai-cockpit
#AI