← Archive

AI Daily — June 25, 2026

2026-06-25

openai.com

OpenAI and Broadcom unveil LLM-optimized inference chip

OpenAI and Broadcom have jointly introduced 'Jalapeño', a custom ASIC designed specifically for LLM inference workloads. The chip targets improved throughput, energy efficiency, and scale compared to general-purpose GPU deployments. This marks OpenAI's first publicly disclosed custom silicon effort, signaling a strategic push to reduce dependence on NVIDIA hardware.

deepmind.google

Introducing computer use in Gemini 3.5 Flash

Google DeepMind has added computer use capabilities to Gemini 3.5 Flash, enabling the model to interact with desktop environments by controlling mouse, keyboard, and screen state. This follows Anthropic's earlier computer use release and positions Gemini as a direct competitor in GUI-based agentic tasks. The Flash variant suggests the feature is targeted at latency-sensitive, cost-efficient deployments.

openai.com

How agents are transforming work

OpenAI published a research paper examining how AI agents are changing knowledge work, with findings showing agents completing longer-horizon, multi-step tasks that were previously impractical to automate. The paper quantifies productivity gains across a range of professional roles and task types. It represents one of the more systematic empirical studies from OpenAI on real-world agentic deployment outcomes.

rss.arxiv.org

Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models

This paper challenges a core assumption in mechanistic interpretability: that detecting a behavior in a model's activations implies you can steer it. The authors measure the angular gap between the linear direction that best detects a behavior and the direction that best causes it, finding that for hallucination in Gemma 2-2B-it, the model achieves perfect linear separability (AUC=1.0) in detection yet the detection and control axes are geometrically misaligned. The result implies that probing-based interpretability success does not automatically translate into reliable behavioral intervention.