The Register
·
2 months ago
Amazon's custom semiconductor business exceeded a $20 billion annual run rate and ranks among the top three datacenter chip makers globally, with its Trainium AI training chips and Graviton processors growing over 100 percent year over year. The company has secured major commitments including two gigawatts of capacity from OpenAI, five gigawatts from Anthropic, and deployment agreements with Meta and Uber, with Trainium2 largely sold out and Trainium3 nearly fully subscribed. Amazon's chip business enables AWS to expand its AI services, with Bedrock token processing exceeding all prior years combined and customer spending growing 170 percent quarter over quarter.
Google Research
·
2 months ago
● 2 sources
Google released Empirical Research Assistance (ERA), an AI system that helps scientists generate expert-level empirical software, which has been applied to four real-world problems: generating weekly forecasts for flu, COVID-19, and RSV hospitalizations that match or exceed CDC tools; deriving six general solutions for the unsolved cosmology problem of gravitational energy radiation from cosmic strings; developing a physics-guided neural network to map column-averaged CO2 every 10 minutes using existing weather satellites instead of the 16-day revisit cycle of specialized sensors; and discovering interpretable neural circuits in zebrafish by combining ERA with structural wiring information. Google has been submitting prospective forecasts to CDC challenges since November 2024 and performing at or near the top of public leaderboards for flu and COVID-19. The work demonstrates how AI-backed systems can solve open problems, democratize access to computational modeling, and extract additional value from existing observational instruments across diverse scientific fields.
Amazon Science
·
2 months ago
Researchers demonstrated that machine learning models trained on sensitive data can leak information about their training datasets through inference attacks, federated learning gradient reconstruction, and shared global models. Specific defenses include differential privacy (which adds calibrated noise during training, achieving 78% accuracy at privacy budget ε=1.5 versus 90% without privacy protection) and secure multiparty computation (which successfully prevented reconstruction of training samples in experiments). Organizations deploying models on private financial, healthcare, or proprietary data must now implement these technical mitigations to comply with regulations like HIPAA and GDPR.
The Register
·
2 months ago
A security engineer created a fake Wikipedia entry claiming he won a nonexistent German card game championship and seeded a domain to support it, then showed multiple AI chatbots confidently cited his fabricated victory when asked. The attack required only a $12 domain registration, a single Wikipedia edit, and about 20 minutes of effort. The experiment reveals that AI systems relying on web search cannot distinguish legitimate sources from newly registered domains, and that poisoned training data could persist in future models even after cleanup, representing a significant security risk especially if exploited to control AI agent actions.
Hugging Face Blog
·
2 months ago
IBM released Granite 4.1, a family of dense language models in three sizes (3B, 8B, and 30B parameters) trained on approximately 15 trillion tokens across a five-phase pipeline that progressively shifts from general web data to curated, high-quality content. The 8B instruct model matches or exceeds the performance of the previous 32B Granite 4.0-H-Small model despite using fewer parameters and a simpler architecture. The models undergo supervised fine-tuning on 4.1 million quality-controlled samples and multi-stage reinforcement learning to improve performance on math, coding, and instruction-following tasks.
OpenAI Blog
·
2 months ago
OpenAI is expanding its Stargate data center project to provide computational capacity for advancing artificial intelligence capabilities. The company is adding new data center capacity, though specific investment amounts or timeline details are not provided in this announcement. This infrastructure expansion aims to support OpenAI's continued development of larger AI models and meet increasing computational demands from users and applications.
The Register
·
2 months ago
Amazon's internal StoreGen team uses AI coding tools like Kiro to accelerate development, but requires human review of all output before deployment, contradicting AWS's keynote rhetoric about AI as magic. The team found that spec-driven development reduces but does not eliminate AI hallucinations and unexpected behavior, with engineers spending less than 30 percent of their time on core coding tasks. Amazon maintains mandatory human approval for all mutating actions including code deployment and document publishing, reflecting a measured approach to agentic AI rather than full automation.
The Register
·
2 months ago
ZTE deployed a 5G-A private network solution at the 2026 Jiangsu Football City League opening ceremony to support two R050 humanoid traffic police robots managing crowds and traffic. The network achieved uplink speeds of 2 Gbps, downlink speeds of 6 Gbps, latency below 10 ms, and 99.99% reliability. The deployment demonstrates a scalable model for using robots in security and traffic management at large sporting events.
Allen Institute (AI2)
·
2 months ago
Researchers at AI2 released MolmoPoint, a vision-language model that points by selecting directly from visual inputs rather than generating text coordinates, achieving state-of-the-art results on pointing and tracking benchmarks among open models. The model demonstrated significant improvements in training efficiency and end-task performance, particularly for high-resolution images and cluttered user interfaces. AI2 also released MolmoWeb, a suite of web agents that navigate websites and complete tasks using only screenshots and mouse/keyboard controls, outperforming comparable open models and matching larger proprietary systems like GPT-4o.
OpenAI Blog
·
2 months ago
OpenAI released a five-part action plan for improving cybersecurity as AI systems become more prevalent. The plan emphasizes democratizing AI-powered cyber defense tools and protecting critical infrastructure from emerging threats. Organizations are expected to adopt these AI-powered security measures to address vulnerabilities that traditional defenses may miss.
.png)
Together AI
·
2 months ago
DeepSeek-V4 Pro, a 1.6-trillion-parameter mixture-of-experts model, is now available on Together AI with a 512K-token context window for long-context reasoning tasks. The model costs $2.10 per million input tokens, $0.20 per million cached input tokens, and $4.40 per million output tokens, with cached input representing a 90% cost reduction for reused context. Teams can choose between three reasoning modes (Non-Think, Think High, Think Max) and deploy via serverless inference or dedicated infrastructure, enabling applications like code agents, document analysis, and research synthesis without requiring self-hosted deployment.
Hugging Face Blog
·
2 months ago
DeepInfra, a serverless AI inference platform, is now integrated as a supported provider on the Hugging Face Hub, allowing developers to run models directly through Hugging Face's infrastructure. The platform offers access to over 100 models including DeepSeek V4, Kimi-K.2.6, and GLM-5.1, with initial support for conversational and text-generation tasks launching immediately. Developers can now use DeepInfra through Hugging Face's Python and JavaScript SDKs, web UI, and agent frameworks without additional setup, with billing passed through directly at provider rates or routed through Hugging Face accounts.
