Physicist Giorgio Parisi and Francesco Zamponi used Anthropic's Claude to prove a decade-old mathematical conjecture about jamming, marking a milestone in AI-assisted scientific discovery.
Intelligence
Curated AI news and interesting developments
AI chip startup Etched emerges from four years of stealth with a working transformer ASIC, $800M in funding, and over $1B in customer contracts — positioning itself as a credible Nvidia alternative for inference.
Meta FAIR's latest non-invasive brain-to-text model achieves 39% word error rate using MEG scans, narrowing the gap with surgical implants through data scaling.
NVIDIA's novel diffusion language model architecture separates context and denoising into two towers, achieving 2.42× throughput while retaining 98.7% of autoregressive quality.
OpenAI unveils the GPT-5.6 family with three tiers — Sol, Terra, and Luna — alongside a limited preview emphasizing coding, cybersecurity, and long-horizon agentic tasks.
Claude Sonnet 5 bridges the gap between Sonnet and Opus, delivering near-Opus-level agentic performance at a significantly lower price point with enhanced safety.
Claude Fable 5, Anthropic's most capable public model, returns globally on July 1 after an 18-day suspension triggered by US government export controls over cybersecurity concerns.
Apple's biggest AI overhaul yet: a rebuilt Siri with deep contextual awareness, a family of five Google-collaborated Foundation Models, and on-device AI for every developer.
The latest Gemma 4 model variant runs complex multimodal AI workloads locally on standard laptops with just 16GB of RAM, challenging the cloud-first assumption for AI deployment.
NVIDIA's customizable multimodal safety model addresses the growing need for scalable, compliant AI safety mechanisms across text, image, and audio modalities.
Anthropic released Claude Tag, an always-on AI agent that lives inside Slack channels as a shared team member, enabling task delegation, ambient monitoring, and autonomous async work.
Google DeepMind introduced Gemma 4 12B, a 12-billion-parameter multimodal model with an encoder-free architecture that runs locally on devices with 16GB of VRAM, released under Apache 2.0.
John Jumper, co-creator of AlphaFold and 2024 Nobel Prize winner in Chemistry, announced he is leaving Google DeepMind after nearly nine years to join rival AI lab Anthropic.
Norway announced sweeping age-based restrictions on generative AI in classrooms, effectively banning AI for children aged 6-13 and limiting use for older students, citing declining test scores.
SpaceX agreed to acquire Anysphere, the startup behind AI coding assistant Cursor, for $60 billion in an all-stock deal to boost xAI's competitive position against OpenAI and Anthropic.
DeepReinforce launched Ornith-1.0, a family of four open-source coding models ranging from 9B to 397B parameters that write their own training scaffold during reinforcement learning — all released under the MIT license on Hugging Face.
NVIDIA released NeMo AutoModel, an open-source tool that accelerates MoE model fine-tuning by 3.4-3.7x and reduces GPU memory by 29-32% with just a single import line change to existing Hugging Face Transformers v5 code.
Alibaba's Qwen team released the first language world models capable of simulating seven diverse agentic environments, outperforming GPT and Claude on environment simulation while enabling cheaper, controllable agent training.