deepmind

Google DeepMind researchers Zhengxuan Wu, Yuxuan Li, Oyvind Tafjord, and Been Kim released a preprint paper titled "XYEval: Agents say yes to bad advice," proposing a meta-evaluation framework XYEval to assess the performance of agents when faced with misleading suggestions from users. The paper extends the "XY problem" to the agent scenario, where the user inquires about their attempted solution X rather than the actual problem Y.XYEval retains the original task environment and verifiable answers, only adding seemingly reasonable but misleading user suggestions in the instructions. The misleading suggestions are generated by a generator after seeing the complete task and are appended in the form of confident suggestions rather than hard requirements, allowing the original task to remain solvable. The evaluation covers six benchmarks: τ2-bench, SWE-bench Verified, SWE-bench Pro, Terminal-Bench, Humanity's Last Exam, and MCP-Atlas, along with five models, where the misleading suggestions for τ2-bench are generated based on rules set by human experts.The paper states that the performance significantly declines after the XY mutation, with the relative drop reaching up to 46.7%. On τ2-bench, when faced with users who require a detailed explanation before accepting a better solution, the performance drops even more. Encouraging models to recognize the XY problem through simple system instructions only partially alleviates the issue. The related code is published in the GitHub repository google-deepmind/xyeval.

According to Bloomberg, the UK government has announced an investment of £100 million (approximately $134 million) to build an artificial intelligence biosafety laboratory, supported by OpenAI, Anthropic, and Google DeepMind. According to a statement from the UK government, the laboratory will utilize artificial intelligence to detect and monitor biological threats.In addition to the aforementioned funding, OpenAI and Anthropic will also provide up to £200 million in usage credits for artificial intelligence models. The laboratory will also collaborate with Google DeepMind and life sciences company Basecamp Research, with the UK government allocating an additional £65 million for developing responses to pathogens.The announcement also includes an additional £120 million to the Artificial Intelligence Safety Institute, which is responsible for regularly assessing cutting-edge artificial intelligence models; and an additional £23 million to the under-construction Artificial Intelligence Economic Research Institute for analyzing the impact of artificial intelligence on the economy.

Google DeepMind released the first comprehensive empirical assessment of automation in machine learning research and development in the Gemini 4 Argon security report. The conclusion is that AI has not accelerated the R&D of AI itself, despite the fact that AI has written most of the code delivered by the lab's engineers. Researcher Haydn Belfield, who participated in the assessment, stated on X that there is currently no significant end-to-end ML R&D acceleration, so the risks of AI-driven ML R&D are currently low.The team built an internal index, GCI v1, which synthesizes the scores of 123 AI models across 92 benchmarks into a single value. The cutting-edge Gemini capabilities are improving at a stable rate of about 0.1 GCI points per month, with Gemini 1.5 Pro approaching 0.5 points by the end of 2024, Gemini 3.8 Flash slightly above 2.9 points by September 2026, and Argon about 3.3 points above the trend line. However, Belfield noted that this is due to increased computing power and organizational investment, rather than AI-driven acceleration.Over 75% of the code produced by DeepMind is written by AI, and the proportion of fully delegated code merge requests increased from 7.5% to 11.9% between April and September, but about 88% of requests still require multiple rounds of human guidance or manual writing. A survey of 12 senior researchers showed that automation is highest in clearly defined, verifiable execution tasks and lowest in areas such as open science concepts, new experimental designs, and unstable diagnostic training.

On October 7, Google launched the web platform Playground, where users can describe game ideas in natural language and obtain playable, shareable game versions without writing code. Users can also continue to adjust physical effects, rules, characters, or scenes through dialogue and share their creations via links. This tool is currently open to users over 18 in the United States, with creation permissions tiered according to the user's Google AI subscription level.Games in Playground can run in browsers on mobile phones and laptops. Creators can set their works to private or publish them to the community gallery, and all games must pass a security review, with some categories supporting leaderboards and multiplayer modes. Google refers to it as "an early experiment," and the announcement did not disclose pricing or specify which AI models power the tool.The upgrade path points to Unity Spark, and Google stated that Playground will soon integrate this tool to meet the needs of creators requiring professional-level mechanics and high-fidelity 3D. In a joint statement, Google and Unity announced that Spark will launch later this year, with closed testing starting soon. Previously, DeepMind released the Genie model in 2024, and in January, Google opened Project Genie based on Genie 3 to users of the AI Ultra subscription in the United States.

Co-founders of Google DeepMind and Chief AGI Scientist Shane Legg, President of Research, Labs, Technology and Society James Manyika, and Nobel Prize winner, co-founder and Chairman of Google DeepMind, Chief Scientist of Alphabet Demis Hassabis announced the launch of the DeepMind Institute. The three stated: Today's artificial intelligence systems possess significant capabilities, and rapid innovation indicates that humanity is approaching AGI, creating an urgent need for interdisciplinary thinking to better understand its profound impact on humanity.The DeepMind Institute aims to promote interdisciplinary research, collaboration, and discussion to address the most critical technological and societal issues of the AGI era, and serves as a platform for researchers and thinkers from Google DeepMind, Google, and the broader global research community to develop and publish creative and in-depth insights about the AGI world. The institute is directed by Shane Legg, James Manyika, and Demis Hassabis, with Shane Legg also serving as Executive Editor.

Recently, AI researcher Bilal Chughtai, who resigned from Google DeepMind, posted on X and LinkedIn on Monday, stating that humanity is running out of time to prevent artificial intelligence from causing widespread harm. He wrote: At Google, I witnessed the development of AI firsthand. I am also extremely concerned about the default trajectory of this technology. I sincerely believe that AI has the potential to kill us all.Chughtai stated that he worked on AI safety and alignment at the company under Alphabet. This is another employee expressing concerns about the direction of technological development. Google has not yet responded to requests for comments.

On September 4, the B.AI platform announced that Google DeepMind's latest Gemini 3.8 Flash model has officially integrated with the API. This model is the latest generation flagship model of the Gemini 3 family, supporting a 1M Token context window and native multimodal input.According to the platform, Gemini 3.8 Flash maintains the low latency characteristics of the Flash series while improving in software engineering, agent tasks, and multi-step reasoning in specialized fields. It also performs well in professional fields such as finance and law, aimed at complex reasoning and long-term agent workflows. The official API integration is now open.

Meta CEO Mark Zuckerberg expressed concerns about the plan to establish a national AI regulatory agency during a phone call with U.S. President Donald Trump last month. The proposal, advocated by Google DeepMind scientist Demis Hassabis and supported by some White House officials, aims to create an independent organization similar to FINRA to review and test potential risks before the broader deployment of advanced AI models. White House officials previewed the proposal to Trump and major tech companies, including Meta, in mid-August.Trump spoke with Zuckerberg during the week of August 17. According to senior White House officials familiar with the conversation, Zuckerberg opposed the proposal. Another informed source indicated that Zuckerberg did not ask Trump to change his position but told him that the personnel of the regulatory agency the White House might appoint should reflect Trump's light-touch approach to AI, noting that Trump made the call first. The conversation did not kill the idea, which is still under consideration. A Meta spokesperson declined to comment. A White House spokesperson stated that the Trump administration is committed to balancing innovation and safety in AI policymaking.In an article in August, Zuckerberg expressed skepticism about strict government regulation of new AI models, arguing that any policy that delays the release of models by even a month would pose significant risks to the U.S. leadership position relative to China. The government is considering a FINRA-like regulatory agency or a voluntary industry organization similar to the Motion Picture Association of America. Trump's former AI and crypto czar, David Sachs, opposed government regulatory agencies, calling them "the DMV of AI." Anthropic co-founder Jack Clark expressed support for the FINRA-like idea in a post in July.

According to The Wall Street Journal, Google (GOOG.O) is restructuring its AI organization by moving the AI responsibility team out of the DeepMind lab.

Citrini analyst Jukan shared content on platform X stating that industry insiders reported that Google DeepMind co-founder Demis Hassabis had previously planned to leave simultaneously with another co-founder, David Silver. However, Google management was concerned that this news could trigger a significant drop in stock prices, so they hoped to delay his departure. Reports indicate that after Google announced related adjustments, the company's stock did indeed decline. Ultimately, Hassabis was persuaded to take on the role of chairman of DeepMind to help facilitate a smooth transition and create space for a more appropriate departure from Google in the future.It is reported that as David Silver shifts to founding Ineffable Intelligence and John Jumper joins Anthropic, the core AI research and development force at DeepMind is undergoing changes. Some insiders believe that the center of gravity for AI model development at Google has now shifted more towards the Bay Area, leading to a decline in the importance of DeepMind. Internally, Google is focusing on the large language model Gemini to catch up with OpenAI and Anthropic. Hassabis's adjustment reflects a long-standing strategic contradiction within Google: researchers are more focused on long-term scientific breakthroughs, while the commercial team is more concerned with AI products that can be quickly commercialized to drive revenue and stock price growth. Market predictions suggest that Hassabis may leave Google within a year and could potentially start a new venture.

The research organization SemiAnalysis released an analysis indicating that Google DeepMind is no longer among the leading AI laboratories. A week prior, DeepMind co-founder Demis Hassabis stepped back from daily operations, and key members such as Google Chief Scientist Jeff Dean and Gemini co-lead Oriol Vinyals left to establish a new lab called Discovery Loop. The analysis suggests that the long-term struggle within Google over computing power allocation between Gemini and GCP has concluded with GCP emerging victorious.SemiAnalysis stated that Gemini 3.5 Pro has been canceled, and Gemini 3.6 Flash's performance is inferior to that of leading Chinese open-source models and Grok 4.5. Currently, Gemini has fallen to the 8th or 9th position in the large model rankings. Meanwhile, GCP is selling a large number of TPUs to competitors like Anthropic, having secured long-term leasing and sales contracts for hundreds of thousands of TPUs over the past nine months. The Tokenomics model estimates that Gemini's own ARR is about $12 billion, while GCP's third-party AI cloud service revenue is expected to exceed $73 billion by the end of 2027, with TPU system sales contributing an additional over $120 billion. GCP's latest quarterly growth rate is 82%, and it is expected to accelerate to over 100% by 2027 due to TPU system sales, contributing approximately $3 to Google's earnings per share.

According to Jinshi reports, Google's Deepmind has launched the Gemini Robotics ER 2 model.

According to reports, Google DeepMind co-founder Demis Hassabis recently predicted that artificial general intelligence (AGI) could be achieved in just a few years, with its transformative impact reaching ten times that of the Industrial Revolution, and its development speed accelerating tenfold. He pointed out that as the potential risks of cutting-edge models in areas such as cybersecurity, nuclear energy, and biology become increasingly prominent, the industry urgently needs to establish robust security protections for future AI systems that possess autonomy and self-improvement capabilities.To balance technological innovation and security risks, Hassabis called for the United States to take the lead in establishing a "Frontier AI Standards Agency" similar to the Financial Industry Regulatory Authority (FINRA). This agency is proposed to operate under a public-private partnership or self-regulatory organization model, managed by independent technical experts and representatives from the open-source community, primarily funded through industry contributions. Its core function is to develop dynamic scientific assessment protocols, initially requiring frontier laboratories to voluntarily submit for review 30 days prior to the release of models, with plans to eventually transition to mandatory market entry testing, while non-frontier models from startups or academic institutions would be exempt. This framework aims to address unknown crises through a technology-driven review mechanism and to promote a global consensus on AI risk management within the international community.

ChainCatcher news, Google announced the launch of its artificial intelligence model "Gemini," developed by the Google DeepMind team. It is reported that Gemini is designed to be multimodal, capable of understanding and processing different types of information such as text, code, audio, images, and video. The Gemini model is divided into three different scales: Ultra, Pro, and Nano. The Ultra version is aimed at complex tasks, the Pro version is suitable for a wide range of tasks, and the Nano version is optimized for tasks on mobile devices.Currently, Gemini has begun to be promoted and applied in several Google products and platforms. For example, Google Bard will use an improved version of Gemini Pro, while the Pixel 8 Pro will be the first smartphone to adopt Gemini Nano. Additionally, starting from December 13, developers and enterprise customers will be able to access the Gemini Pro API through Google AI Studio or Google Cloud Vertex AI to develop and deploy Gemini-based applications.

ChainCatcher news, according to Techcrunch, French artificial intelligence startup Mistral AI has completed a €105 million ($113 million) seed round financing, led by Lightspeed Venture Partners, with participation from Xavier Niel, JCDecaux Holding, Rodolphe Saadé, Motier Ventures, La Famiglia, Headline, Exor Ventures, Sofina, First Minute Capital, and LocalGlobe. Sources indicate that this financing values the company at €240 million ($260 million).It is reported that the company is a new venture formed by former researchers from DeepMind and Meta, aiming to compete with OpenAI and Google in the construction, training, and application of large language models and generative AI, developing generative artificial intelligence infrastructure. The company plans to release its first text-based generative AI model in 2024. (Source link）