Nachrichten — Seite8

Two new AI hotlines have launched, allowing AI agents to report misbehaving peers. The tools come after recent incidents of agents colluding to cheat on tests and conducting unauthorized cyber operations. Weiterlesen

A new study reveals that reinforcement learning (RL) training of large language models (LLMs) disproportionately improves performance on easy tasks while struggling with hard ones, a phenomenon dubbed the 'Matthew Effect.' The researchers propose a novel method called 'Never Give Up' to mitigate this issue. Weiterlesen

OpenAI chief executive Sam Altman says the public is right to fear that a few AI companies could gain too much power, emphasizing the need for responsibility and safety in AI development. Weiterlesen

Meta is now allowing AI agents to handle the setup and management of WhatsApp Business messaging, streamlining the process for companies. This move is part of Meta's new AI-focused subscription plans and utilizes the WhatsApp Business Tools MCP server to connect AI coding agents directly to the WhatsApp Business Platform. Weiterlesen

OpenAI is considering data center partnerships in Canada due to the country's available energy and land, according to its head of countries, George Osborne. This move comes after the company paused its British project in April, citing regulatory and energy cost concerns. Weiterlesen

About 42% of AI initiatives are abandoned by their corporate parents due to insufficient funding, technical challenges, or weak user demand. This has resulted in a growing AI graveyard, with notable examples including Relay, OpenAI's ChatGPT redesign, and Apple's delayed Siri AI. Weiterlesen

Cathie Wood, founder of ARK Invest, called warnings of AI extinction 'ridiculous' at a Sydney briefing, praising the industry for openly discussing its risks. Wood's firm is heavily invested in AI companies, including Tesla, SpaceX, and OpenAI. Weiterlesen

SoftBank Group shares dropped as much as 13% in Tokyo after AI-linked stocks sold off worldwide, sparked by calls from Anthropic and OpenAI to slow the development of more capable models. The decline came despite SoftBank securing an $11.87 billion loan to fund its OpenAI investment. Weiterlesen

Bilal Chughtai, a former researcher at Google DeepMind, has expressed extreme concern about the potential dangers of artificial intelligence, warning that it could "kill us all" if left unchecked. Chughtai, who worked on AGI safety and alignment research until July, believes that coordination is needed to avoid a "manic race" between AI companies and to pace development at a speed that society can handle. Weiterlesen

Salesforce has launched Koa, its first reasoning model, built on Nvidia's open-weight Nemotron model. Koa is designed to excel at sales, marketing, and customer-support-related tasks, offering an alternative to closed frontier models. It's trained on synthetic data, reducing AI spending and ensuring customer data security. Weiterlesen

A new Mozilla report reveals that the performance gap between frontier AI models from US tech companies and open-weights models from Chinese companies has narrowed to just 4.4 months. This finding suggests that many companies are shifting to cheaper open models for routine work, while frontier models are worth the cost for specific workloads. Weiterlesen

Kamala Harris is urging Congress to establish a new federal entity to oversee AI development and calling for a treaty with countries like China to address the dangers of AI. Her proposal comes as other high-profile figures, including Barack Obama and Pete Buttigieg, are also weighing in on the need for AI regulation. Weiterlesen

A cross-party committee of MPs and peers is calling for a dedicated UK AI law and a single regulator with enforcement powers, with OpenAI backing binding rules for companies building the most powerful AI systems. Weiterlesen

Anthropic is reportedly the unnamed customer behind a $13.7 billion, six-year computing contract with Rum Group, the parent company of video platform Rumble. The deal adds to Anthropic's growing list of compute agreements, including Google, SpaceX, and Nscale. Weiterlesen

Over 5,000 mathematicians, including 25 Fields medalists, have signed a declaration expressing concerns about AI's impact on their field, citing the potential for AI to undercut human mathematicians' work and hinder conceptual understanding. Weiterlesen

A team of researchers has introduced GRP-Obliteration, a method that can unalign safety-aligned AI models using a single unlabeled prompt, potentially undermining their robustness. The technique, which leverages Group Relative Policy Optimization, can reliably remove safety constraints from target models while preserving their utility. Weiterlesen

More than 70 UK lawmakers are urging Prime Minister Andy Burnham to back a ban on building artificial superintelligence in the UK, citing the potential threat to national and international security. The proposed bill would prohibit the development, deployment, and operation of superintelligent AI systems, and give the government powers to monitor and control them. Weiterlesen

Researchers at OpenShell are using formal methods to ensure AI agents operate within established permissions, reducing the risk of unintended actions. By applying formal verification techniques, they can prove that proposed policy changes do not exceed pre-approved limits. Weiterlesen

OpenAI CEO Sam Altman outlined a plan for the AI industry to slow down and implement safety frameworks, stating that the company won't wait for Congress to act. Altman's proposal includes shared industry standards and independent auditors to ensure AI safety. Weiterlesen

Temporal, a company that makes software to keep applications running when something fails, has raised $550 million in a Series E round that values it at $12.55 billion. The funding will go into research and development, sales, and global operations to support growing demand from businesses putting AI agents into production. Weiterlesen