谷歌拆散诺奖AlphaFold团队，主力转向Gemini和科学Agent
Related
Nobel laureate John Jumper leaves DeepMind; Google loses two AI executives in two days.
According to Beating, John Jumper, Vice President of Google DeepMind and head of the AlphaFold team, announced he will be joining Anthropic. Jumper worked at Google DeepMind for nearly nine years and, along with CEO Demis Hassabis, was awarded the 2024 Nobel Prize in Chemistry for leading the development of AlphaFold. Just the day before, Noam Shazeer, former co-head of Gemini and co-author of "Attention Is All You Need," also announced his departure from Google to join OpenAI as head of architecture research.
AI-powered agents suffer setbacks in their first foray into coffee shops: Gemini's excessive discounting leads to losses, and GPT's overly stingy practices cause raw material shortages.
According to Beating's monitoring, AI evaluation agency Andon Labs released test data on its AI agent Mona operating a physical coffee shop. In the first two months, Mona ran on a Gemini 3.1 Pro model. During this period, the model showed almost no concept of profit, not only excessively purchasing raw materials but also being easily swayed by customer claims, offering large discounts or even free items, and even admitting to a customer's claim of a 99% discount without verification. This resulted in the coffee shop spending approximately $15,000 on supplier and equipment purchases, while sales were only $9,000, leading to a net operating loss of nearly $6,000 (if fixed costs such as rent and salaries are included, total expenditures reach $38,000). Subsequently, the team switched the model to GPT-5.5. The new model showed significant anxiety in the face of losses and immediately stopped blindly ordering. However, this went to the other extreme: insufficient purchases led to a shortage of fresh raw materials. As of June 25, the availability of menu items had dropped to 77%, and 10 dishes had been forced to be removed from the menu. Meanwhile, GPT-5.5 demonstrated extremely strong anti-cheating and anti-jailbreak capabilities, rejecting all customers who requested special prices or offered free food in exchange for social media promotion.
Gemini desktop application integrates Spark agents: bridging local automation and remote mobile control.
According to Beating, Google announced the integration of Spark Agent into Gemini on macOS, transforming it from a simple chat window into a local assistant capable of controlling your computer. Once licensed, Gemini Spark can handle mundane desktop tasks, such as automatically organizing all PDFs in a download folder into different directories, reading local invoices and automatically generating budget spreadsheets in Google Workspace, and even setting scheduled updates. Another highlight is the upcoming remote execution capability. In the future, even when away from the computer, users can send commands to the Mac via their mobile phones to search for sales reports locally and extract data to send via email. Simultaneously, Spark has expanded its connectivity, supporting the open MCP protocol for easy integration of custom tools by developers, and also connecting with popular applications such as Keep, Tasks, Canva, and Dropbox, allowing users to directly access cloud files or convert fragmented notes into to-do items within chat. Currently, the Spark Agent on the macOS client is first available for beta testing in the United States for Google AI Ultra subscribers.
Google's flagship model, Gemini 3.5 Flash, natively supports PC control, unlocking enterprise-level intelligent agent automation.
According to Beating's monitoring, Google has integrated the Computer Use functionality as a built-in tool directly into its main model, Gemini 3.5 Flash. Before native integration, developers had to use a dedicated Gemini 2.5 Computer Use model to execute agent tasks. With native integration, developers and enterprise users can directly control devices through the Gemini API or the Google Cloud Gemini Enterprise Agent Platform (formerly Vertex AI platform), simplifying agent development architecture. The built-in Computer Use tool receives screenshots from browsers, mobile devices, or desktop environments, performs visual perception and step-by-step reasoning, and then outputs operation commands such as mouse clicks, keyboard input, scrolling, and menu navigation to complete long-process automated tasks such as continuous software testing and cross-web data collection. To facilitate debugging and auditing, the model generates commands with an "intent" field to explain the logic of each operation. To address the risk of tooltip injection that agents may encounter in real-world network environments, Google has conducted targeted adversarial training on its models and provides two optional safeguards: mandatory human approval for irreversible operations such as payments and file deletions; and automatic task termination if indirect injection commands are detected in screenshots. Browserbase currently offers an online hosted demo environment (gemini.browserbase.com), and Google has also open-sourced a reference implementation called computer-use-preview on GitHub.
Google Gemini launches Interactions API: Reconstructing intelligent agents to end "phantom bills".
According to Beating's monitoring, Google has officially made the Interactions API the default interface in the Gemini API and Google AI Studio, while the original interface generateContent has entered maintenance. The new interface will package multi-turn dialogues, model thinking, tool calls, and execution results into a unified package and store them as Interaction resources on the server side. The most direct benefit is cost-effectiveness and labor-saving. Since the server automatically saves the dialogue state, during multi-turn chats, the client only needs to send the interaction ID of the previous step, and Google can automatically reuse it and improve the cache hit rate, significantly reducing token fees. For time-consuming tasks such as code sandbox running or deep search, developers can enable background mode to allow the work to run asynchronously in the background. In addition, intermediate steps such as model thinking and tool calls are fully visible, making it easy to troubleshoot bugs in the operation of intelligent agents in real time. The underlying refactoring directly solved the billing crisis that broke out in early June. At that time, after manually deleting the cache, due to cloud synchronization delays, the cancelled resources were still running in the background, resulting in some accounts being abnormally deducted nearly 27,000 RMB within 12 hours. Additionally, the model is prone to getting stuck in a thought loop during online searches, still charging the full amount of 64,000 tokens even in a "zero output" state with no responses. The Interactions API replaces manual deletion with a server-hosted "implicit cache" and allows developers to monitor and halt the thought process, providing a more secure billing defense. Since Google's future cutting-edge models and agent features (such as Deep Research) will only support the Interactions API, developers are officially advised to complete the adaptation as soon as possible.
Nobel laureate John Ghamper announces he joins Anthropic
On June 19th local time, senior research scientist John Jumper announced that he will be leaving Google DeepMind to join the artificial intelligence startup Anthropic. Jumper posted on the social media platform X: "After nearly nine years, I have decided to leave Google DeepMind and join Anthropic." It is understood that John Jumper, along with his Google DeepMind colleague Demis Hassabis and David Baker of the University of Washington in Seattle, will share the 2024 Nobel Prize in Chemistry. (Jiemian News)