AI Digest

AI21 Labs Releases Jamba 1.6 with Extended 256K Context Window
AI21 Labs unveiled Jamba 1.6, the latest iteration of its hybrid SSM-Transformer architecture model, now supporting a 256,000-token context window. The model demonstrates significant improvements in long-context recall and reasoning tasks, positioning it as a direct competitor to Anthropic's Claude and Google's Gemini. Early benchmarks show it outperforming GPT-4 on several summarization and multi-document analysis benchmarks.

EU Parliament Approves Final Language for AI Liability Directive
The European Parliament voted to approve the final text of the AI Liability Directive, which establishes a rebuttable presumption of causality in cases involving high-risk AI systems. The directive shifts the burden of proof to companies to demonstrate their AI was not at fault in causing harm, a major win for consumer advocacy groups. The regulation is expected to take effect in member states by early 2028.

Google DeepMind Introduces AlphaFold 3 for Drug Discovery
Google DeepMind released AlphaFold 3, capable of predicting the structure of protein complexes with small molecules, nucleic acids, and post-translational modifications. The model achieves a 70% accuracy rate on protein-ligand interactions, a key bottleneck in computational drug design. Pharmaceutical partners including Novartis and Pfizer will begin testing the model in their R&D pipelines immediately.

Meta Open-Sources Large-Scale Synthetic Data Generator for Robotics
Meta published "RoboGen," a framework for generating unlimited synthetic training data for robotic manipulation tasks using physics simulations. The tool allows researchers to create millions of diverse task variations without manual labeling, drastically reducing the cost of training generalist robot policies. Early tests show a 40% improvement in zero-shot transfer to real-world environments.

Nvidia Announces Blackwell Ultra GPU with 2x Memory Bandwidth
Nvidia revealed the Blackwell Ultra GPU at its GTC Europe event, featuring 2.4 TB/s memory bandwidth and a new FP6 compute mode for more efficient large model inference. The chip is designed specifically for serving trillion-parameter models, with a claimed 4x improvement in tokens-per-second per watt over the H100. Mass production is slated for Q4 2026, with cloud availability expected in early 2027.

US Commerce Department Proposes Mandatory Safety Audits for Frontier Models
The US Department of Commerce released a draft rule requiring developers of AI models exceeding 10^26 FLOPs of training compute to submit annual third-party safety audits. The rule would apply to companies like OpenAI, Anthropic, and Google, mandating red-teaming results, capability evaluation reports, and incident logs. Public comment period runs through July 2026.

Runway Gen-4 Achieves Real-Time Video Generation at 30 FPS
Runway released Gen-4, its latest video generation model capable of producing 1080p video at 30 frames per second with less than 200ms latency on a single A100. The model supports text-to-video and image-to-video modes, and introduces "temporal consistency" controls that allow users to lock specific objects or characters across frames. The update is immediately available to Pro subscribers.

OpenAI Launches ChatGPT "Canvas" for Collaborative Editing
OpenAI introduced Canvas, a new interface within ChatGPT that allows users to edit, annotate, and version control long-form text and code outputs side-by-side with the model. The feature supports real-time collaboration with multiple users and integrates with GitHub and Google Docs for export. Canvas is rolling out to ChatGPT Plus and Team subscribers over the next week.

Editor's take: Today's news underscores a shift from raw model capability to deployment infrastructure and regulation. The EU's liability directive and the US's proposed safety audits signal that governments are moving beyond voluntary commitments, while Meta's synthetic data generator and Runway's real-time video show the industry is doubling down on practical, production-ready tools. The AI landscape is maturing—less about "can we build it" and more about "how do we trust and deploy it safely."