Anthropic has released its new 'Mythos-class' frontier AI models, Claude Fable 5 and Claude Mythos 5, which feature a 1 million token context window and demonstrate state-of-the-art capabilities in coding, vision, and scientific research. However, the launch has sparked controversy due to aggressive safety filters that trigger fallbacks to older models, a mandatory 30-day data retention policy on AWS Bedrock, and undocumented performance throttling designed to restrict competing LLM development.
Google has released Gemini 3.5 Live Translate, a new streaming speech-to-speech model supporting over 70 languages that preserves a speaker's original pitch and intonation. The technology is being integrated into Google Meet, the Google Translate app, and developer platforms, while also being piloted by ride-hailing giant Grab. To combat misinformation, Google is embedding imperceptible SynthID watermarks into all audio generated by the model.
Cohere co-founder Nick Frosst has launched North Mini Code, an open-source 30-billion parameter Mixture-of-Experts model with 3 billion active parameters designed for local execution. Released under the Apache 2.0 license on Hugging Face, the model is optimized for agentic software engineering tasks and supports MLX in VS Code.
Prominent AI figures, including AI2's Nathan Lambert and researcher Jeremy Howard, have publicly criticized Anthropic for its increasingly antagonistic stance toward open-weight models and open AI research. Critics argue that Anthropic's strategy of restricting access to its most advanced models while using them internally for frontier research exacerbates power imbalances and contradicts safety goals. This pushback highlights a growing ideological divide between corporate AI developers and proponents of democratized, open-source AI development.
An autonomous drone boat, the Saronic Corsair, successfully rescued a downed U.S. Army Apache helicopter crew, marking the first unmanned military sea recovery. This event demonstrates the critical operational capabilities and real-world application of AI-powered autonomous systems in military operations.
China is embarking on a five-year plan to invest $295 billion in national data center construction to boost its AI capabilities, highlighting the country's strategic commitment to advancing its AI sector despite potential chip constraints. This massive investment aims to strengthen China's foundational infrastructure for artificial intelligence development.
A Munich regional court has ruled that Google is directly responsible for false claims made by its AI-generated search overviews, rejecting the defense that users should fact-check the results. The court determined that AI overviews generate independent statements rather than merely linking to third-party content, meaning traditional search engine liability shields do not apply. Furthermore, the ruling classified AI-generated opinions as business activities rather than protected free speech, establishing a landmark precedent that could impact other AI providers like OpenAI, Anthropic, and Perplexity.
New research reveals that low-bit KV cache quantization, a common technique for reducing LLM inference memory, can silently destroy safety alignment in instruction-tuned models, a problem not detected by standard perplexity metrics. The study introduces Per-Channel Reduction (PCR), a training-free diagnostic and mitigation protocol that recovers up to 97% of lost alignment, addressing critical vulnerabilities in production LLM serving.
Google DeepMind has introduced Gemma 4 12B, a highly efficient multimodal AI model designed to run locally on consumer laptops with 16GB of RAM using a novel encoder-free architecture. Alongside the model, DeepMind released the Gemma Skills Repository, a library of tools and resources designed to help developers build advanced AI agents using the Gemma ecosystem.
Amazon Web Services has demonstrated a new generative AI system designed to automate the insurance First Notice of Loss (FNOL) intake process by processing multimodal evidence like photos, videos, and audio. Powered by the open-source Strands Agents SDK, Amazon Bedrock, and the new Amazon Nova Act SDK, the system uses AI agents to navigate web portals and apply codified domain expertise to structure and tag claim data. This automation reduces repetitive screen-based tasks for adjusters, allowing them to focus on high-value decision-making.
Tesla has secured regulatory approval to deploy its Full Self-Driving (FSD) Supervised software in Denmark, marking a significant expansion of its autonomous driving technology into a new European market. This decision has also sparked discussions and questions within the community regarding a potential subsequent rollout in Portugal.
Noam Brown, co-creator of OpenAI's o1, urges benchmark developers to plot LLM performance against test-time compute, revealing that GPT-5.5 outperforms GPT-5.4 when token budgets are equalized. This discussion is crucial for accurately evaluating and comparing AI model capabilities.
Boris Cherny, a lead on Anthropic's Claude Code team, has introduced an experimental update that enables nested subagent support, capped at five levels. This feature aims to enhance the management of context limits within complex AI workflows, improving the model's ability to handle intricate tasks.
Victor Taelin, founder of Higher Order Company, announced that his AI agent Fable achieved a significant 17.7x speedup on the HVM5 evaluator benchmark, surpassing previous results from GPT-5.5 agent swarms. This demonstrates a notable advancement in AI agent performance and optimization.
OpenAI's Chief Futurist Joshua Achiam argues that Anthropic focuses on building protective AI to oversee humanity, contrasting with OpenAI's approach of directly distributing AI tools to users. This highlights a significant philosophical divergence between two leading AI companies on the future role and deployment of artificial intelligence.