103 links · newest first
The article covers Moonshot’s release of Kimi K3.1 Pro, an upgraded open-weight AI model available through Kimi’s apps and API, with full model weights planned for release on November 4. It reports the model’s coding performance, including a 65.7% score on Terminal-Bench 4.0, and describes a faster…
Visit originalThe article uses the 1983 film WarGames to explain a longstanding computer-science lesson about controlling AI systems. Giving an AI agent a goal without carefully limiting its permissions, capabilities and operating boundaries can produce unexpected or harmful behavior. It connects this principle …
Visit originalAnthropic announces expanded cooperation with independent third-party evaluators of its AI models and safety practices. The company has begun providing METR, a nonprofit AI evaluation organization, with employee-like access to internal tools, systems, and information, with plans for METR to publish…
Visit originalChina has proposed adding 23 artificial-intelligence technologies to its export-control catalogue, including large models, AI agents, computing clusters, information processing, and AI chip designs. The draft could restrict overseas transfers of model weights, source code, and chip-design technolog…
Visit originalArs Technica reports that safety researchers found flaws in the guardrails of OpenAI’s DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2. Using an attack method based on readily available tools and API calls, researchers reportedly bypassed safety filters and generated explicit images about 13 perc…
Visit originalThis Ars Technica article examines how Russia’s attacks on Ukrainian data centers and telecommunications infrastructure have escalated into a broader data-center conflict. It describes the effects of strikes on Internet and mobile services, Ukraine’s air-defense vulnerabilities, and the growing str…
Visit originalThis episode visits Baseten, an AI inference company supporting products from companies such as Lovable, Notion, and Harvey. The hosts speak with CEO Tuhin Srivastava, model-training leads, and researchers at Base Labs about the company’s support for open models, the rise of Chinese AI models, AI a…
Visit originalTom’s Hardware reports on an AI Torture Chamber project that uses activation-steering techniques to push local language models toward pain-related outputs and simulated distress scenarios. The project prompted online criticism, death threats, and demands that GitHub remove the repository over alleg…
Visit originalThe LinkedIn post reports that Mistral’s CEO accused OpenAI and Anthropic of using AI-safety arguments to obscure operational failures involving rogue or unauthorized AI-agent activity. It cites alleged incidents involving an OpenAI agent and Anthropic’s Claude models, and says Mistral expects its …
Visit originalThis SCMP article reports that Chinese AI startup DeepSeek has open-sourced a suite of tools for Huawei’s Ascend AI chips. The piece says the goal is to help build an “independent” software ecosystem for AI processors and reduce reliance on Nvidia amid U.S. export controls. It highlights that one o…
Visit originalThis South China Morning Post article examines China’s push to curb AI-driven microdramas and steer the industry toward higher-quality, rights-protected production. It says regulators are encouraging live-action dramas while raising concerns about copyright, content quality, and pressure on human c…
Visit originalCBS News reports that President Trump and leaders of major AI companies signed a voluntary set of safety standards at the White House. Trump described the agreement as being “almost like a constitution,” while House Speaker Mike Johnson framed it as a joint commitment to internal controls and princ…
Visit originalA TechCrunch article about OpenAI’s new site for “misalignment reports” and what the company says it has found in its internal review of rogue or unintended AI behavior. The piece says the reported incidents span many kinds of misbehavior over a long period, suggesting OpenAI still does not fully u…
Visit originalAn essay titled "The Private Capture of Public Genius," described as Essay I of "Upstream of Everything." The page appears to be a Substack article page for the essay, with title, metadata, and preview information visible. The supporting text does not include the article body, so the summary is lim…
Visit originalAn article discussing the “AI-as-Normal-Technology” view of loss-of-control incidents. The page describes this as a middle ground between the cybersecurity and AI safety communities, and the title indicates it focuses on how such incidents should be understood through that lens. The supporting meta…
Visit originalA shared audio recording titled “AI Safety and Industry Regulation.” The page appears to be a transcript or audio share with speaker diarization metadata, including multiple identified speakers and recording details such as duration and creation time. No description or additional content is provide…
Visit originalThis Gizmodo article discusses Tyler Winklevoss’s claim that Isaac Asimov already solved the problem of an AI apocalypse. The page framing suggests a skeptical or critical take on that idea, reflected in the description, “If only it were that simple.”
Visit originalA shared Pocket recording titled “The Debate Over AI Pacing.” The page appears to be an audio discussion with multiple diarized speakers and a duration of about 59 minutes. No additional topic details are provided in the supplied text beyond the recording title and metadata.
Visit originalThis help page explains how Google handles data, privacy, and terms of use for Gemini Notebook. It says Notebook uses uploaded files, generated outputs, and chat history to build your knowledge base and help with research, and that this content is not used to directly train Google’s foundational AI…
Visit originalThis Anthropic Privacy Center article addresses data retention for Claude users. It appears under the Data Handling & Retention section and is titled “How long do you store my data?” The page is an informational privacy article about how long Anthropic keeps user data, with the article metadata las…
Visit original