NeuralStackUK Introduces AI Safety Testing Mandate Effective October 2026
The UK government published new AI safety testing rules on July 14 effective October 1. Developers of models above 10^26 FLOPs must submit third-party evaluations. Non-compliance carries fines up to 6 percent of global revenue.
Stanford Researchers Achieve 94 Percent on GPQA with New Architecture
Stanford University researchers published a new architecture on July 11 that scores 94 percent on GPQA. The model uses dynamic sparse attention and requires 60 percent less compute than dense transformers. The paper details training on 2.1 trillion tokens.
Amazon Web Services Launches Bedrock 3 with Agent Builder
AWS released Bedrock 3 on July 14 introducing a visual agent builder and 12 new foundation models. The update supports multi-agent orchestration and native integration with SageMaker. Customers can now deploy production agents in under 30 minutes.
Cloudflare Mitigates July 10 DDoS Attack on 1.2 Million Sites
Cloudflare blocked a record 3.8 Tbps DDoS attack on July 10 targeting 1.2 million customer sites. The assault originated from compromised IoT devices across 140 countries. No customer data was accessed during the six-hour event.
Perplexity AI Raises 1.6 Billion Series E at 18 Billion Valuation
Perplexity AI closed a 1.6 billion Series E round on July 13 at an 18 billion valuation. The round was led by a16z and included new participation from Fidelity. Funds will expand real-time search infrastructure and enterprise API offerings.
Cohere Releases Command R+ 2 with 5M Context Window
Cohere launched Command R+ 2 on July 12 featuring a 5 million token context window and improved multilingual support. The model scores 92 percent on MMLU and processes enterprise documents 40 percent faster than prior versions. It targets legal and financial sectors seeking long-document analysis without external retrieval.
Amazon Announces Q3 2026 Rollout of Custom Trainium3 Chips
Amazon Web Services will deploy Trainium3 chips across all US regions starting September 15. The new silicon delivers 2.8 times higher training throughput than Trainium2 at equivalent power draw. AWS customers can access the chips through EC2 Trn3 instances with per-second billing.
Oracle Reports July 11 Breach Affecting 1.4 Million Cloud Tenants
Oracle confirmed a July 11 intrusion that exposed records of 1.4 million cloud tenants. Attackers accessed customer metadata and limited configuration data through a compromised internal dashboard. The company has isolated affected systems and begun mandatory password resets for all impacted accounts.
Inflection AI Raises 1.9 Billion Series E at 22 Billion Valuation
Inflection AI closed a 1.9 billion dollar Series E round on July 11 at a 22 billion post-money valuation. Microsoft and NVIDIA led the round with participation from existing backers. The capital funds expansion of its personal AI agent platform to 50 million users by year end.
NVIDIA Launches Nemotron-4 120B with 8M Context for Enterprise AI
NVIDIA released Nemotron-4 120B on July 12 featuring native 8 million token context and 94 percent MMLU accuracy. The model runs on Blackwell GPUs with 40 percent lower inference latency than prior versions. Enterprises gain real-time document analysis and agent orchestration capabilities previously limited to research labs.
OpenAI and Microsoft Release Triton 3 Inference Framework
OpenAI and Microsoft launched Triton 3 on July 12 2026 with automatic kernel fusion and 3x faster MoE inference. The open-source framework supports PyTorch 2.7 and CUDA 13. Developers can achieve 420 tokens per second on H200 GPUs for 70B models.
Apple Unveils On-Device 120B Parameter Model for iOS 20
Apple introduced a 120 billion parameter on-device model on July 9 2026 for iOS 20. The model runs entirely on A20 silicon with 4-bit quantization and delivers 78.3 percent on MMLU. It powers new Siri features launching in September.