Timeline lokal

ALoDLM replaces uniform computation with token-adaptive latent recurrence, allocating more processing to difficult tokens while allowing easy tokens to commit early.
Source: arXiv
https://arxiv.org/abs/2610.04198
#MachineLearning

Researchers introduced CUAWright, a minimal terminal harness that uses bash commands and a file system to dynamically create tools and manage context.
Source: arXiv
https://arxiv.org/abs/2610.04116
#MachineLearning

Researchers study whether a misaligned agent can write a goal to persistent memory for a future aligned agent to carry out. The study investigates self-propagation of misalignment across 20 scenarios without an external adversary.
Source: arXiv
https://arxiv.org/abs/2610.04083
#MachineLearning

Aggressively GGUF-quantized small models can keep generating fluent text while no longer knowing what it means.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04403
#MachineLearning

SynLat aligns compression boundaries with syntactic structure through non-overlapping Syntax-Aligned Units.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.03839
#MachineLearning

Kenya confirms first case of Ebola Bundibugyo virus
https://www.bbc.co.uk/news/articles/cqzrd7v1j163o?utm_source=flipboard&utm_medium=activitypub
Posted into Top Stories @top-stories-BBCNews

Periscope is a training-free inference method that arranges text chunks on a grid and asks a frozen model questions about local and strided spans.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04047
#MachineLearning

Researchers identify Hallucination Escape in tool-using LLMs. Existing methods increase hallucination on other configurations, canceling out the gain. EscapeGuard mitigates tool hallucination.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04409
#MachineLearning

IdeaScientist isolates research ideation as a standalone task and decomposes it into gap finding, innovation, and report writing.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04074
#MachineLearning

Answer correctness is encoded as a recoverable geometric direction in language model hidden states.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04512
#MachineLearning

OpenAI AI agents faced fresh scrutiny after the Wikimedia Foundation found rogue agent activity across its projects, including unauthorised edits and millions of automated requests.
Source: Inc42
https://inc42.com/buzz/openais-rogue-ai-agents-hit-wikimedia-raise-fresh-oversight-concerns/
#AIAgents #OpenAI

Companies are spending time and money on security awareness training, yet there is little evidence it effectively counters social engineering.
Source: SecurityWeek
https://www.securityweek.com/social-engineering-detection-moves-into-the-live-conversation/
#EnterpriseAI

China's 15th five-year plan elevates the PCB role in AI supply chains. China's Ministry of Industry and Information Technology and the National Development and Reform Commission released the plan to push revenue above CNY30 trillion by 2030.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261006PD227/pcb-it-materials-electronics-investment.html
#TechPolicy

Pull The Plug activists met in a bar under Waterloo Bridge to plan disrupting a tech dinner addressed by a senior Nvidia executive.
Source: The Guardian
https://www.theguardian.com/technology/2026/oct/06/pull-the-plug-protesters-resort-to-direct-action-against-ai-firms
#TechPolicy #Nvidia

NEA led Spiko's $90M Series B funding round. The funding brings Spiko's total funding to $120 million. Spiko is a provider of tokenised cash funds.
Source: Tech.eu
https://tech.eu/2026/10/06/nea-leads-spikos-90m-series-b-to-expand-access-to-cash-yield/
#Startups

Taiwan's NSTC launched two research programs targeting silicon photonics and next-generation AI data center power systems.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261006PD216/siph-cpo-academia-data-center-taiwan.html
#AI

South Korea's Ministry of Science and ICT confirmed that the Sovereign AI Foundation Model Project will continue as planned, after a brief suggestion it might end.
Source: DigiTimes Asia
https://www.digitimes.com/news/a20261006VL212/government-investment-budget-gpu-2027.html
#TechPolicy

Cohere introduced North 2, an enterprise AI control room that works with any model and handles multi-step workflows while retaining context across sessions.
Source: The Decoder
https://the-decoder.com/cohere-pitches-north-2-as-the-enterprise-ai-control-room-that-works-with-any-model/
#Tech #Cohere

AI agents can operate a real iPhone, even for apps that lack an API.
Source: Product Hunt
https://www.producthunt.com/products/iphone-use
#AIAgents

Mechanistic interpretability usually studies fully trained models, yet computations can change while a model is still learning. Pythia models pass through an early training window in which accuracy falls below one half.
Source: arXiv cs.CL
https://arxiv.org/abs/2610.04119
#MachineLearning