Google introduced its eighth-generation Tensor Processing Units (TPUs), two new chips specifically engineered to meet the computational demands of the "agentic era" of AI. This hardware innovation aims to enhance the performance and scalability of advanced AI systems.
Sources: Latent Space, Simon Willison, TLDR AI, The Neuron
Anthropic recently acknowledged and resolved significant performance degradation across its Claude models, particularly Claude Code, stemming from multiple internal changes including a memory-clearing bug, a reverted reasoning effort change, and a problematic system prompt. Following user feedback, the company deployed fixes, reverted changes, and reset usage limits for all subscribers as an apology. Anthropic has committed to implementing stricter internal testing and enhanced code review to prevent future regressions.
Sources: Hacker News, Keep Bookmarks, Simon Willison
Chinese AI lab DeepSeek has released its V4 models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, featuring a groundbreaking 1-million-token context window and designed for highly efficient AI agentic tasks. These models incorporate novel hybrid attention mechanisms and agent-specific features, demonstrating competitive performance against leading frontier models while offering significantly lower pricing due to major efficiency improvements. DeepSeek-V4-Pro, with 1.6 trillion parameters, is also noted as potentially the largest open-weights model available.
Sources: Hugging Face Blog, Simon Willison
A brief mention of "preview access" to OpenAI Codex with GPT-5.5 indicates that OpenAI is actively developing and testing a new, unreleased iteration of its foundational language model. This suggests an upcoming advancement in OpenAI's AI capabilities, building upon their existing GPT series.
Sources: Keep Bookmarks, Simon Willison
OpenAI launched GPT-5.5, a "worker-class" model now available in ChatGPT and Codex, which outperformed Anthropic's Opus 4.7 in Terminal-Bench 2.0 and GDPval tasks. This release marks a significant competitive advancement for OpenAI in the LLM space.
Sources: The Neuron
Anthropic has reportedly reached a $1 trillion valuation in secondary markets, exceeding OpenAI's valuation. This milestone reflects strong investor confidence and reshapes the financial landscape among leading AI development companies.
Sources: The Neuron
Sony AI's Ace robot has successfully defeated elite human ping-pong professionals under official rules, marking a significant milestone as the first AI to conquer a physical, real-time, adversarial sport.
Sources: The Neuron
The Pentagon has requested a record $54 billion budget for autonomous drone warfare in FY2027, marking the largest commitment to AI-powered military technology in history, raising concerns among experts about deployment doctrine.
Sources: The Neuron
OpenAI has launched ChatGPT for Clinicians, which reportedly surpassed specialty-matched physicians on the new HealthBench Professional benchmark, demonstrating advanced AI capabilities in medical contexts.
Sources: The Neuron
Agility Robotics has deployed its Digit humanoid robots in Schaeffler's automotive parts factory, marking one of the first operational uses of humanoids in industrial settings. These robots perform tasks like ferrying parts, matching the cost of human labor and potentially shifting human workers to supervisory roles, indicating their economic viability for specific factory tasks.
Sources: The Batch
Google has introduced TorchTPU, a new software stack enabling native and efficient PyTorch model execution on its Tensor Processing Units (TPUs). This initiative features an "Eager First" philosophy with a Fused Eager mode for significant performance gains, comprehensive support for distributed training, and leverages XLA/StableHLO for optimization. A 2026 roadmap includes a public GitHub and integrations with key AI ecosystem projects.
Sources: Hacker News
NVIDIA demonstrated a Gemma 4 Vision-Language Assistant (VLA) running efficiently and locally on a Jetson Orin Nano, autonomously utilizing a webcam for visual context through native tool-calling. This advanced multimodal AI deployment, facilitated by `llama.cpp` with CUDA support, reinforces the Jetson platform's capabilities for edge AI and offers simplified deployment via the Jetson AI Lab.
Sources: Hugging Face Blog
OpenAI released GPT-5.5, which achieves comparable intelligence to Claude Opus 4.7 at a quarter of the cost, establishing a new cost/performance frontier. This model offers improved long-horizon execution, stronger computer-use behavior, and enhanced token efficiency.
Sources: Latent Space
Under the leadership of new CTO Mikhail Parakhin, Shopify has achieved nearly 100% employee adoption of AI tools and developed sophisticated internal platforms like Tangle for reproducible ML, Tangent for automated AI research, and SimGym for customer behavior simulation. The company is also integrating non-transformer Liquid AI models and custom systems to manage the evolving challenges of AI-generated code, marking a deep strategic commitment to AI integration.
Sources: Latent Space
OpenAI has launched its new GPT-5.5 language model to ChatGPT subscribers and Codex, with premium API pricing announced to position it as a high-end offering. The company also clarified its support for third-party integrations via the Codex API, a policy shift that followed the hiring of OpenClaw creator Peter Steinberger. Initial reviews highlight GPT-5.5's 'jagged frontier' of capabilities, showing inconsistent performance despite advancements.
Sources: Simon Willison