Meta is arguing that its use of copyrighted books to train its Llama LLM falls under fair use, even when obtaining training data involves BitTorrent uploads of pirated material. This defense is critical as the lawsuit progresses, focusing on whether the downloading and sharing of copyrighted content also qualifies as fair use.
Sources: 1 Covered by Hacker News
A new analysis reveals that LLMs frequently generate syntactically plausible code that is fundamentally incorrect or highly inefficient, sometimes performing 20,000 times slower than human-written code. This highlights a significant gap between the output of LLMs and the requirements for real-world applications, with LLMs prone to over-engineering and creating complex solutions for simple problems.
Sources: 1 Covered by Hacker News
Anthropic has launched Agent Skills as an open standard, utilizing a SKILL.md format for defining AI agent capabilities. This standard has been adopted by key industry players including Microsoft, OpenAI, Atlassian, GitHub, and Cursor, promoting interoperability across different AI tools.
Sources: 1 Covered by Hacker News
OpenAI's GPT-5.4 (xhigh) has regained the top ranking on the Artificial Analysis Intelligence Index, showcasing improvements in physics reasoning and agentic coding. However, this comes at the cost of higher per-token prices and a reported increase in hallucination rates compared to previous versions.
Sources: via Latent Space
Anthropic's AI model, Claude Opus 4.6, identified 22 vulnerabilities in Firefox over two weeks, with 14 classified as high-severity, and developed 'crude' browser exploits in a testing environment. This highlights AI's potential in cybersecurity and raises concerns about its possible malicious use, leading Anthropic to release Claude Code Security in a limited research preview.
Sources: via Anthropic News
Anthropic, OpenAI, and Google are engaged in contract discussions with the Pentagon. Anthropic is strategically positioning itself as a moral and trustworthy AI provider to differentiate itself in a market where top-tier AI models are becoming increasingly commodified.
Sources: via Simon Willison
Sarvam AI has made its 30B and 105B models available as open-source. This initiative aims to foster innovation and collaboration within the open-source AI community by providing broader access to advanced AI models for researchers and developers.
Sources: via Hacker News
Google's SpeciesNet, an open-source AI model that identifies nearly 2,500 animal categories from camera trap photos, has significantly accelerated wildlife monitoring and research. Projects like Snapshot Serengeti, the Humboldt Institute's Red Otus network, and the Wildlife Observatory of Australia are leveraging SpeciesNet to analyze vast datasets, monitor species, and aid conservation efforts.
Sources: via Google AI Blog
Descript is using OpenAI's AI models to enhance its video editing platform with scalable multilingual video dubbing. This integration optimizes translations for both meaning and timing, providing natural-sounding speech and more efficient localization features.
Sources: via OpenAI Blog
Codex Security has launched an AI-powered application security agent that analyzes project context to detect, validate, and patch complex vulnerabilities. The agent aims to provide higher confidence and less noise in security operations.
Sources: via OpenAI Blog