Headlines

1. Anthropic and OpenAI Announce Joint Frontier Safety Pause Ahead of 2027 Models
On Sept. 24, 2026, the two leading labs agreed to halt pretraining of models exceeding 10^26 FLOPs until third-party safety evaluations are passed, a move aimed at pre-empting regulatory action. Both companies will publish progress reports every 60 days. Source: Reuters

2. EU AI Act High-Risk Rules Take Effect, First Major Fines Issued
The European Commission levied €120 million in fines against two unnamed U.S. tech platforms on Sept. 25 for violating new transparency requirements under the AI Act’s high-risk category. This marks the first enforcement action since the regulations fully came into force this month. Source: Financial Times

3. Global AI Compute Coalition Breaks Ground on Geothermal Exascale Cluster in Iceland
A 10-nation partnership on Sept. 25 began construction of a 2.5-exaflop AI supercomputer powered entirely by Iceland’s geothermal energy, designed to serve as open infrastructure for public-interest AI research. The project expects initial operations by mid-2027. Source: AP News

Model Releases

4. Meta Releases Llama 4 Ultra with Open-Weight 2 Trillion MoE Architecture
Meta launched Llama 4 Ultra on Sept. 24, featuring 2 trillion total parameters (300 billion active), 128K context, and native multimodal generation across text, image, and video. The open-weights release is available via Hugging Face and cloud providers. Source: TechCrunch

5. Mistral AI Launches Codestrider 2 for Autonomous Software Engineering
The Paris-based startup released Codestrider 2 on Sept. 25, a specialized coding LLM with 512K context and an integrated agentic loop that can independently debug, test, and refactor entire repositories. It reportedly sets a new state of the art on SWE-bench Verified 2.0. Source: The Verge

Industry

6. Nvidia Unveils Blackwell Ultra B300, Doubling Inference Throughput
At a Sept. 24 data center event, Nvidia announced the B300 “Blackwell Ultra” GPU delivering 2 peta-FLOPs of FP4 compute and 8 TB/s memory bandwidth, targeting next-generation AI agents and real-time multimodal workloads. Shipments are scheduled for Q1 2027. Source: Reuters

Research Highlights

7. DeepMind’s AlphaFold 4 Maps Dynamic Protein Assemblies at Atomic Resolution
In a Nature paper published Sept. 25, DeepMind detailed AlphaFold 4, which predicts not only static structures but also conformational states and protein-ligand interactions at atomic resolution. The model is expected to accelerate drug discovery and protein design by capturing motion and binding. Source: Nature

8. New Multimodal Scaling Laws Predict Emergent Abilities Across Modalities
A joint Stanford and Meta AI study released on arXiv on Sept. 25 shows that cross-modal loss follows a unified power-law across text, image, and audio, with emergent reasoning abilities appearing after 10 billion trained tokens. The paper proposes a single metric for multimodal model scaling. Source: arXiv preprint

Compiled from multiple media reports.