OpenAI previewed GPT-5.6 Sol, its next-generation model with significant improvements in coding, science, and cybersecurity, paired with its most advanced safety stack. The release comes amid U.S. government regulatory restrictions.
Paul Meade, Apple's vice president in charge of Vision Pro, is reportedly joining OpenAI's hardware team. This move reflects OpenAI's growing interest in developing its own hardware devices.
OpenAI restricted GPT-5.6 rollout at the U.S. government's request, but expressed concern about making this a permanent norm. The measure reflects growing regulatory pressure on advanced AI models.
A 4M-parameter model achieves 98.6% accuracy on symbolic math without explicit math knowledge, suggesting LLMs perform structured pattern completion. This finding challenges whether mathematical "reasoning" in AI is true reasoning or pattern prediction.
OpenAI announced Jalapeño, its custom inference chip built with Broadcom, joining Google, Apple, and SpaceX in building proprietary hardware. The move aims to diversify suppliers and reduce single-vendor dependency costs.
The Trump administration requested OpenAI to launch GPT-5.6 in limited form to select partners rather than the general public due to safety concerns. This marks the first direct regulatory intervention on an AI model release.
The U.S. government will vet users accessing GPT-5.6, setting a precedent for state control over advanced model distribution. This mechanism marks a fundamental shift in how AI models are regulated.
OpenAI confirmed it will launch GPT-5.6 in limited preview format in response to safety concerns from the Trump administration. CEO Sam Altman told employees the strategy will be gradual rather than an immediate open release.
OpenAI previewed GPT-5.6 Sol, its new model with significant improvements in coding, science, and cybersecurity. The release includes enhancements to its safety stack, marking an important advancement in overall capabilities.
Patronus AI, founded by former Meta AI researchers, raised $50M to build "digital worlds" for stress-testing AI agents. The funding reflects growing demand for agent validation tools in production environments.
OpenAI and Broadcom unveiled Jalapeño, a custom chip designed specifically for LLM inference. This initiative aims to improve performance, efficiency, and scalability of production AI systems.
The company bets that millions of hours of gameplay can train AI agents closer to human intuition. This approach could revolutionize how we develop systems that make complex decisions in dynamic environments.
Google added computer interaction capabilities to Gemini 3.5 Flash, allowing the model to directly control computers and automate tasks through graphical interfaces. This feature significantly expands use cases for AI agents.
Un-0, a new image generation system, demonstrates how the technology can replicate conventional AI systems with radically improved efficiency. If confirmed, this could transform the economics of training large models.
Anthropic accused Alibaba of illicitly extracting Claude's capabilities through unauthorized reverse engineering. The dispute raises important questions about model security and IP protection in AI.
OpenAI and Broadcom unveiled Jalapeño, a custom AI chip specifically designed to improve performance and efficiency in LLM inference. This development matters for developers looking to optimize costs and latency in production deployments.
OpenAI revealed its first AI processor specifically designed for LLM inference, developed in partnership with Broadcom. The Jalapeño ASIC chip aims to optimize cost and performance for AI servers, marking OpenAI's entry into the semiconductor market.
Qualcomm completed its acquisition of Modular, a promising AI chip software startup, for nearly $4 billion. This deal reflects the accelerating race to control AI infrastructure and hardware optimization.
Google introduced 'computer use' capability in Gemini 3.5 Flash, enabling the model to directly interact with user interfaces. This feature opens new possibilities for automating tasks that require navigation and application control.
Anthropic launched Claude Tag, enabling teams to collaborate with Claude directly in Slack. 65% of Anthropic's product team's code is already created using this feature, demonstrating its potential as a team workflow tool.
Subscribe to receive the daily 5AI brief every evening, around 7:30 PM Chile time.
English edition.