AI news digest — July 10, 2026
10 items, each with its source.
Meta Releases Llama 4: The First Open-Source AGI Candidate
Meta has open-sourced Llama 4, a massive 2-trillion parameter multimodal model that matches the performance of top proprietary models across all major benchmarks. The release includes full weights and training data, sparking intense debate over the safety of democratizing such powerful AI systems.
ai.meta.comOpenAI Unveils GPT-5.5 with Real-Time Continuous Learning
OpenAI has officially launched GPT-5.5, introducing a novel continuous learning architecture that updates its weights in real-time based on user interactions. This breakthrough significantly reduces hallucination rates and allows the model to adapt to new information without requiring massive retraining runs.
openai.comDeepMind Proposes Liquid State Space Models to Replace Transformers
Google DeepMind published a groundbreaking paper introducing Liquid State Space Models (LSSMs), an architecture that outperforms Transformers in both efficiency and reasoning capabilities. The new approach achieves O(1) inference time while maintaining perfect recall over a 10-million token context window.
arxiv.orgEU AI Act Enters Full Enforcement Phase for Foundation Models
The European Union has officially begun enforcing the strictest tier of the AI Act for general-purpose AI models. Companies failing to comply with the new transparency and risk assessment mandates now face fines of up to 7% of their global annual turnover.
commission.europa.euAlphaFold 4 Successfully Models Complex Protein-Drug Interactions
Isomorphic Labs and Google DeepMind have jointly published research on AlphaFold 4, which can now accurately predict dynamic protein-drug interactions in real-time. This advancement is expected to reduce the early-stage drug discovery timeline from years to mere weeks.
nature.comNvidia Announces Rubin R100 GPU with Photonic Interconnects
Nvidia has unveiled the Rubin R100, its next-generation AI accelerator featuring fully integrated silicon photonics for chip-to-chip communication. The new architecture promises a 10x increase in training speeds for trillion-parameter models while cutting energy consumption in half.
nvidianews.nvidia.comFigure Deploys 10,000 Autonomous Humanoids in BMW Plants
Robotics startup Figure announced the successful deployment of 10,000 Figure 03 humanoid robots across BMW's global manufacturing facilities. The robots are fully autonomous, utilizing a new vision-language-action model to perform complex assembly tasks alongside human workers.
figure.aiApple Integrates On-Device LLMs into Vision Pro 2
Apple announced the integration of its proprietary 30-billion parameter on-device LLM into the newly released Vision Pro 2. The update enables real-time, zero-latency environmental understanding and conversational AI without relying on cloud processing, setting a new standard for spatial computing.
apple.comNuclear AI Startup FissionML Raises $2 Billion Series A
FissionML, a startup dedicated to building small modular reactors specifically for powering AI data centers, has secured a $2 billion Series A funding round. The investment, led by Andreessen Horowitz, highlights the growing energy crisis in the AI industry and the push for sustainable compute power.
techcrunch.comCoalition Launches Cryptographic Watermarking Standard for AI Media
A coalition including Microsoft, Adobe, and the BBC has released a universal cryptographic watermarking standard for AI-generated media. The protocol embeds unalterable metadata into images and videos at the hardware level, providing a robust solution against the proliferation of deepfakes.
c2pa.orgFeed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.