The Pentagon blacklisted Anthropic for refusing to lift safety guardrails for military applications, citing concerns over mass surveillance and autonomous weapons. Shortly after, OpenAI announced its own Pentagon deal, claiming similar safeguards as Anthropic.
Sources: 3 Covered by The Code, The Neuron, The Rundown AI
Researchers found that LLMs (OpenAI, Google, Anthropic) substantially increase novice accuracy in biosecurity-relevant tasks, raising concerns about the accessibility of dangerous knowledge. The study highlights the potential risks associated with the increasing capabilities of AI in sensitive domains.
Sources: via Import AI
Apple has launched a new iPad Air, integrating the M4 chip which significantly boosts performance and AI processing with a faster Neural Engine and increased memory bandwidth. The M4 chip's 16-core Neural Engine is three times faster than its M1 predecessor, enabling advanced on-device AI.
Sources: via Hacker News
/e/OS is an open-source mobile operating system and ecosystem designed to be free from Google services, offering a privacy-enhanced environment with its own apps and online services. It replaces Google services with alternatives like Murena Find and microG.
Sources: via Hacker News
AMD's Am386 CPU was delayed until March 1991 due to extensive legal battles initiated by Intel, who sought to prevent other companies from producing x86-compatible processors. Intel pursued an aggressive legal strategy, spending $100 million over eight years against AMD and other x86-related companies.
Sources: via Hacker News
The libxml2-ee project has proactively addressed critical security vulnerabilities, including use-after-free (UAF) in relaxng, type confusion in c14n, and an infinite loop in xmlCtxtParseContent. This "Enterprise Edition" fork of libxml2 also boasts up to 10x faster parsing with SIMD acceleration.
Sources: via Hacker News
Researchers introduce CoMind, a multi-agent system designed to leverage collective knowledge for machine learning engineering. It achieved a 36% medal rate on past Kaggle competitions, demonstrating its effectiveness in machine learning tasks.
Sources: via arXiv AI
Researchers have developed DataMind, a new framework for building generalist data-analytic agents, which includes a data synthesis recipe and training strategy. Their DataMind-14B model achieved state-of-the-art performance, outperforming GPT-5 in certain data analysis tasks.
Sources: via arXiv AI
A new benchmark, CMT-Benchmark, demonstrates that even frontier large language models like GPT5 struggle significantly with expert-level problems in condensed matter theory. The benchmark comprises 50 problems, with LLMs solving only 30%.
Sources: via arXiv AI