OpenAI Unveils Jalapeño Chip, Noam Shazeer Joins OpenAI, and New AI Science Breakthroughs
OpenAI and Broadcom have introduced the Jalapeño inference chip, while Noam Shazeer has departed Google for OpenAI. Additionally, new AI-driven scientific advancements in immunology and antibody design have emerged, alongside updates on Fable 5 and Google's Gemini 3.5 Flash.
Podcast В· 2 min
OpenAI and Broadcom Unveil Jalapeño Inference Chip
OpenAI and Broadcom have officially unveiled Jalapeño, a custom-built AI inference chip designed to run large language models like ChatGPT more efficiently. Developed in just nine months, the ASIC chip is optimized for real-time model inference rather than training, aiming to reduce OpenAI's heavy reliance on general-purpose Nvidia GPUs. Early testing indicates superior performance-per-watt metrics compared to current state-of-the-art hardware. The chip is already powering the GPT-5.3-Codex-Spark model in OpenAI's labs, with a full deployment scheduled for late 2026, including a significant purchase by Microsoft for 40% of the initial batch. This move represents a major step toward vertical integration, allowing OpenAI to optimize its infrastructure specifically for its own models, potentially lowering operational costs and improving speed.
Noam Shazeer Leaves Google for OpenAI
Noam Shazeer, a co-lead of Google's Gemini model development, has left the company to join OpenAI. This marks Shazeer's second departure from Google, having previously left to found Character.AI before returning to Google in a $2.7 billion deal less than two years ago. His move to OpenAI is a significant talent acquisition for the company, given his foundational role in developing large-scale AI architectures. This departure continues the trend of high-profile AI researchers moving between major labs, highlighting the intense competition for top-tier talent in the frontier AI space.
Google Researchers Join Anthropic
Google DeepMind researchers Jonas Adler, Alexander Prtizel, and Arthur Conmy have announced their departure from Google to join Anthropic. This move is part of a broader wave of talent migration within the AI industry, as researchers seek roles at labs focused on specific safety and scaling paradigms. The loss of these high-profile staff members comes as Google continues to iterate on its Gemini model lineup and agentic capabilities, while Anthropic continues to aggressively expand its research and engineering teams.
Google Engineer Fired After Viral Workspace CLI Tool
Justin Poehnelt, a former Google engineer, reports he was fired shortly after creating and releasing the Google Workspace CLI, an open-source tool designed to allow AI agents to interact with Gmail, Drive, and other Workspace apps via command line. The tool gained significant traction on Hacker News and GitHub, drawing praise from the AI builder community for its agent-native design. The firing has sparked criticism within the tech community, particularly as Google Cloud Next had announced an official Workspace CLI just two days prior. Observers note the irony of the situation, suggesting that internal bureaucracy may be stifling the very innovation that the broader AI ecosystem is actively rewarding.
GPT-5 Pro Solves Immunology Mystery
OpenAI's GPT-5 Pro has been utilized by immunologist Derya Unutmaz to resolve a three-year-old research mystery regarding how T cells—the immune cells responsible for fighting cancer and viruses—specialize. By analyzing complex biological data, the model provided insights that opened new avenues for cancer and autoimmune research. This application demonstrates the growing utility of frontier models in accelerating scientific discovery by processing vast datasets that exceed human analytical capacity in traditional research timelines.
Nabla Bio's JAM-2 Designs Drug-Quality Antibodies
Nabla Bio's JAM-2 has become the first AI model to successfully design drug-quality antibodies directly from a computer. The model achieved binding success rates that match or exceed traditional lab-based discovery methods, specifically against difficult-to-target proteins associated with cancer. This milestone suggests a shift in drug discovery, where computational design could significantly reduce the time and cost required to identify viable therapeutic candidates.
Nvidia Releases BioNeMo Agent Toolkit
Nvidia has launched the BioNeMo Agent Toolkit, a framework designed to enable AI agents to function as junior scientists. The toolkit allows agents to read research papers, generate hypotheses, write code, and iterate on experimental results without constant human intervention. The tool is specifically aimed at life sciences and drug discovery teams, intending to automate repetitive early-stage research tasks and accelerate the development of new therapeutics.
Anthropic and OpenAI Join $500M Intercept Fund
Anthropic, the OpenAI Foundation, and Stripe are among the donors backing Intercept, a new $500 million nonprofit organization. The fund is dedicated to financing the development of respiratory-virus prevention tools, including shots, sprays, and pills, as well as indoor air-cleaning technologies. The initiative aims to address the estimated $600 billion in global productivity losses caused by routine respiratory infections, focusing on bridging the funding gap for early-stage research that traditional pharmaceutical companies often avoid.
Fable 5 Comeback Hints in Claude Code
Recent updates to Claude Code v2.1.190 contain strings referencing Fable 5, suggesting that the model may be returning to availability. The references hint at the game being included in subscriptions with specific usage limits, though nothing has been officially confirmed by Anthropic. This follows a period where Fable and Mythos were taken offline in compliance with U.S. orders, and industry observers are closely watching for signs of the model's return to the frontier.
Google DeepMind Adds Computer-Use to Gemini 3.5 Flash
Google DeepMind has integrated computer-use functionality directly into Gemini 3.5 Flash. This update allows the model to power agents capable of navigating browsers, desktop applications, and operating systems. This development is part of a broader industry push to create agents that can perform multi-step tasks across various software environments, moving beyond simple chat interfaces.
xAI's Grok Traffic Driven by NSFW Content
A report from The Information indicates that a substantial portion of traffic to xAI's Grok is driven by NSFW (Not Safe For Work) content. Employees at xAI have reportedly stated that well over half of the model's traffic comes from explicit generations. This highlights the challenges and market realities faced by AI labs when balancing open-ended model capabilities with content moderation and brand safety.