Gemini 3.5 Flash Adds Computer Use Capabilities — With a Strong Focus on Agent Safety
On June 24, Google DeepMind announced that it has added computer use capabilities to Gemini 3.5 Flash. The feature enables AI agents to actually perform operations on the user's computer, and it is notable for employing adversarial training as a countermeasure against prompt injection.
Two optional features are also being offered for enterprises: one requires user confirmation for sensitive or irreversible operations, and the other automatically halts a task when indirect prompt injection is detected. Google recommends a "defense in depth" approach, calling for it to be combined with sandboxing, human-in-the-loop verification, and strict access controls.
As agentic AI moves toward practical deployment, this is a release that puts safety front and center.
Qualcomm to Acquire AI Software Company Modular for Approximately $4 Billion
According to Reuters, Qualcomm has revealed plans to acquire Modular, an AI software startup, for approximately $4 billion. Modular is known for its technology that optimizes the execution environment for AI/ML workloads, with particular strength in improving AI inference performance.
Qualcomm is a major supplier of mobile SoCs, and the acquisition appears to be aimed at strengthening its AI software layer — seemingly with the intent of boosting its AI competitiveness on both the hardware and software fronts.
OpenAI and Broadcom Jointly Develop an LLM-Optimized Inference Chip
OpenAI has announced that it developed a dedicated LLM inference chip in collaboration with Broadcom. Internally codenamed "Jalapeno", the chip has the following characteristics:
Performance per watt that significantly exceeds the current state of the art
Designed from the ground up to support both current and future LLMs
From design to mass production in just 9 months (leveraging OpenAI's own models)
Plans for gigawatt-scale data center deployment
While OpenAI has so far delivered models and products, it is now steering toward a full-stack platform that includes the chip layer. At this point, however, the chip appears to be aimed at data centers rather than consumers.
SK Hynix Eyes $29 Billion U.S. Listing — Semiconductor Market Booms on AI Demand
SK Hynix, a major South Korean memory semiconductor maker, has announced plans to raise up to $29 billion in the U.S. market through ADRs (American Depositary Receipts). The move comes against the backdrop of surging demand for HBM (high-bandwidth memory) driven by the AI boom.
SK Hynix is a leading HBM supplier for AI chip vendors including NVIDIA, and the listing is presumably intended to raise funds for expanding production capacity and investing in next-generation technologies. It suggests that AI infrastructure demand is reshaping the very structure of the semiconductor industry.
Agility Robotics Plans $2.5 Billion IPO via SPAC
Agility Robotics, a humanoid robot developer, has announced plans to go public through a SPAC (special purpose acquisition company). The deal values the company at approximately $2.5 billion, with $620 million in funding expected.
Spun off from Oregon State University in 2015, the company develops a bipedal robot called "Digit". In the humanoid robotics space, Figure AI, Tesla Optimus, and others are also drawing attention, and with the race toward commercialization intensifying, Agility aims to secure a financial edge by going public.
White House Distances Itself from Anthropic CEO Dario Amodei
According to WIRED, the Trump administration's White House has been scaling back its relationship with Anthropic co-founder and CEO Dario Amodei. At high-level meetings, co-founder Tom Brown has been representing the company in Amodei's place.
A White House official reportedly described Amodei as "eccentric", suggesting that the balance of power between industry and government over AI policy is shifting. The move is once again drawing attention to how AI company leadership influences policymaking.
Sakana AI — The Push to Build Frontier AI in Japan
Sakana AI, a Japanese AI startup, had its initiative titled "Building Frontier AI in Japan" featured on Hacker News. The company has drawn attention for its evolutionary AI research and efficiency work on small-scale models, and it is now showing a serious commitment to developing frontier-class AI models as well.
As an example of a Japan-born AI company taking on global competition, interest is rising within Japan's AI community.
---